Research Summary

Chat Ts&Cs

A Survey of 20 Standard Contracts for Generative AI Services

This page summarises the results of a detailed survey of 20 standard contracts for chat-based generative AI services from 13 leading providers, covering the terms and conditions that govern both commercial and consumer use of the services for customers in the UK and across Europe.

Download the full paper free from SSRN →
Providers and Services Surveyed
#ProviderHQService(s)Intended Use
1.Alibaba CloudChinaQwen ChatConsumer
2.AmazonUSAmazon QCommercial
3.AnthropicUSClaudeConsumer & Commercial
4.BaiduChinaErnie BotConsumer
5.GoogleUSGeminiConsumer & Commercial
Gemini WorkspaceCommercial
6.DeepSeekChinaDeepSeekConsumer & Commercial
7.MetaUSMeta AIConsumer & Commercial
8.MicrosoftUSCopilotConsumer
Microsoft 365 CopilotCommercial
9.MistralFranceLe ChatConsumer & Commercial
10.OpenAIUSChatGPTConsumer & Commercial
ChatGPT EnterpriseCommercial
11.Perplexity AIUSPerplexityConsumer & Commercial
12.ProtonSwitzerlandLumoConsumer & Commercial
13.xAIUSGrokConsumer & Commercial
* 13 providers offering 20 services in total.
1. Contracting Party and Dispute Resolution
1.1 Contracting Party
  • US Providers: The ToS of seven of the eight providers who are headquartered in the US specified as a contracting party either a subsidiary in Ireland or a company registered in a US State. For example, customers of Anthropic's Claude; Google's Gemini Workspace; Microsoft's Copilot (basic) and Microsoft 365 Copilot; and OpenAI's ChatGPT Enterprise contracted with an Irish subsidiary.1Claude Consumer ToS; Claude Commercial ToS; Copilot MSA, S.10; Gemini GCE; Copilot MCA, under "Definitions"; ChatGPT Enterprise OpenAI SA, S.17. Grok customers contracted directly with X.AI LLC, registered in the State of Nevada, while Perplexity customers contracted with Perplexity AI Inc., registered in the State of Delaware.2Grok ToS-C; Grok ToS-E; Perplexity ToS-C; Perplexity ToS-E. By contrast, all European customers of Amazon Q contracted with a subsidiary in Luxembourg.3Amazon Q CA, "Table: Contracting party".
  • Customer location: In some cases, the contracting party differed depending on the customer's location. For Google's Gemini (basic); Meta's MetaAI; and OpenAI's ChatGPT (basic), customers in the European Economic Area ('EEA') contracted with an Irish subsidiary, while customers in the UK contracted directly with a US company registered in the State of Delaware.4ChatGPT Europe ToU, under "Who we are"; Gemini UK ToS under "Service provider"; Gemini EU ToS under "Service provider"; Meta AI UK ToS; Meta AI EU ToS.
  • European and Chinese Providers: Both of the providers headquartered in Europe specified a company registered in the country of the provider's headquarters. Thus, the ToS for Le Chat and Lumo specified parties in France and Switzerland, respectively.5Le Chat ToS (consumer); Le Chat ToS (commercial); Lumo ToS. The ToS for DeepSeek similarly specified a contracting party in China.6DeepSeek ToU. By contrast, the ToS for Alibaba Cloud's Qwen Chat and Baidu's Ernie Bot identified subsidiary companies in Singapore (not China) as the contracting party.7Qwen Chat ToS; Ernie Bot ToS.
1.2 Choice of Law
Choice of law for commercial use8This table shows the number of ToS with a choice of law clause that applied to commercial use of the service per each jurisdiction. The total count for the table is 14. It does not include Alibaba Qwen Chat or Ernie Bot, since the services were for non-commercial use only, nor does it include the consumer ToS for Le Chat, Claude, Copilot (basic), Grok, and Perplexity, since these only apply to consumer use of the service. Further, MetaAI appears twice, since it had separate choices of law for commercial use in the UK (California) and in Europe (Ireland).
#JurisdictionCountServices
1.Ireland4ChatGPT Enterprise, Claude (commercial), Meta AI (EU), Microsoft 365 Copilot
2.California4ChatGPT (basic), Gemini Workspace, Meta AI (UK), Perplexity (commercial)
3.China1DeepSeek
4.France1Le Chat (commercial)
5.Luxembourg1Amazon Q
6.Tennessee1Grok (commercial)
7.Switzerland1Lumo
8.Customer's local law1Gemini (basic)
Choice of law for consumer use9This table shows the number of providers with a choice of law clause that applied to consumer use of the service per each jurisdiction. The total count for the table is 12, as it does not include Amazon Q, Microsoft 365 Copilot, Gemini Workspace, or Chat-GPT Enterprise since these services were aimed at non-consumer use. Nor does it include the commercial ToS for Claude, Grok, Le Chat, or Perplexity since these only applied to commercial use.
#JurisdictionCountServices
1.Customer's local law3Gemini (basic), ChatGPT (basic), Meta AI
2.Singapore2Qwen Chat, Ernie Bot
3.China1DeepSeek
4.England1Claude (consumer)
5.France1Le Chat (consumer)
6.Ireland1Copilot (basic)
7.Switzerland1Lumo
8.Texas1Grok (consumer)
9.California1Perplexity (consumer)
Details
  • Overview: All the ToS we surveyed asserted that any disputes arising under the contract would be governed by the laws of a particular jurisdiction. Overall, Irish and California law were popular choices, especially for clauses that applied to business customers.
  • Common approach. The choice of law often matched the location of the provider's contracting party. For example, the ToS for Amazon Q chose the laws of Luxembourg;10Amazon Q CA, S.11(4), Table "Governing Laws and Governing Courts". while the ToS for both Microsoft's Copilot (basic) and Microsoft 365 Copilot chose Irish law.11Copilot MSA, S.10; Copilot MCA, under "Miscellaneous"(i). Similarly, the ToS for DeepSeek chose the laws of China, while those of Alibaba Cloud's Qwen Chat and Baidu's Ernie Bot chose the laws of Singapore.12DeepSeek ToU, S.9(1); Ernie Bot ToS, S.9(1); Qwen Chat ToS, S.XII. Further, the ToS for Mistral's Le Chat and Proton's Lumo chose the laws of France and Switzerland, respectively.13Le Chat ToS (consumer), S.11; Le Chat ToS (commercial), S.14(12)(3); Lumo ToS, S.10.
  • Perplexity and Grok. Other providers had choice of law clauses that did not match the location of the provider's contracting party. For example, the Perplexity ToS had a contracting party in Delaware, but chose California law.14Perplexity ToS-C, S.11(7); Perplexity ToS-E, S.10(1). Further, the Grok ToS also had a contracting party in Delaware, but the ToS (consumer) chose the laws of the State of Texas, while the ToS (commercial) chose those of the State of Tennessee.15Grok ToS-C, under "Governing law", "Europe Specific Terms"; Grok ToS-E, S.14(6).
  • Consumer v. commercial use. In some cases, the choice of law differed between consumer and commercial use. For example, Anthropic's Claude had different terms for UK consumer and for UK commercial use. The ToS for Claude (consumer) were governed by English law, while those for Claude (commercial) were governed by Irish law.16Claude Consumer ToS, S.13; Claude Commercial ToS, S.M(7). The situation was more complicated for Meta AI, since the choice of law depended on both the customer's location (in the UK or elsewhere in Europe) and their status (as a consumer or a business). The Meta AI ToS (consumer) chose the local laws of the consumer's residence for UK and EU consumers.17Meta AIs UK ToS, S.7; Meta AIs EU ToS, S.7. By contrast, the MetaAI ToS (commercial) chose California law for UK customers, but Irish law for EU customers.18Meta AIs UK ToS, S.7; Meta AIs EU ToS, S.7. This applied to use "in any other capacity", which included "business or commercial" purposes.
  • ChatGPT. The situation was also complicated for ChatGPT. As noted above, OpenAI offered separate services: ChatGPT (basic) and ChatGPT Enterprise.19See e.g. OpenAI, "Introducing ChatGPT Enterprise", https://openai.com/index/introducing-chatgpt-enterprise/. The ToS for ChatGPT (basic) chose local law for consumer use; but California law for "commercial or business use".20Chat-GPT Europe ToU, under "General Terms" and "Business use of the Services addendum". By contrast, the ToS for ChatGPT Enterprise chose Irish law for all enterprise customers in the EEA and the UK.21Chat-GPT Enterprise OpenAI SA S.16(3), 17 under "Governing Laws". Thus OpenAI's ToS contained three different choice of law clauses: they chose the consumer's local laws for consumer use of ChatGPT (basic); California law for commercial use of ChatGPT (basic); and Irish law for ChatGPT Enterprise users. The ToS for Google's Gemini took a slightly simpler approach. The ToS for Google Gemini (basic) chose the customer's local laws for all customers (both consumers and businesses).22Gemini UK ToS and Gemini EU ToS under "Settling disputes, governing law, and courts". This term stood out from the others we surveyed, since it applied the customer's local law to customers who used the service for commercial purposes. In other ToS, this approach was typically restricted to consumers. By contrast, the ToS for Google Gemini Workspace chose California law for all European customers.23Gemini RM, S.I.
  • Analysis. For an analysis of the choice of law clauses under European consumer protection law, download our full paper free from SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698.
1.3 Choice of Forum and Arbitration
Choice of forum for commercial use24This table shows the number of ToS with a choice of forum clause that applied to commercial use of the service per each jurisdiction. The total count for the table is 15. It does not include Alibaba Qwen Chat, since the service was for non-commercial use only. Nor does it include the consumer ToS for Claude, Copilot (basic), Grok, Le Chat, or Perplexity, since these applied only to consumer use. MetaAI appears twice, since it had separate choices of law for commercial use in the UK (California) and in Europe (Ireland).
#JurisdictionCountServices
1.Arbitration4ChatGPT Enterprise, Claude (commercial), Ernie Bot, Gemini Workspace
2.State of California3Meta AI (UK), ChatGPT (basic), Perplexity (commercial)
3.Ireland2Microsoft 365 Copilot, Meta AI (EU)
4.Customer's local courts1Gemini (basic)
5.China1DeepSeek
6.France1Le Chat (commercial)
7.Luxembourg1Amazon Q
8.Switzerland1Lumo
9.Texas1Grok (commercial)
Choice of forum for consumer use25This table shows the number of providers with a choice of law clause that applied to consumer use of the service per each jurisdiction. The total count for the table is 12. It does not include Amazon Q, Chat-GPT Enterprise, Microsoft 365 Copilot, or Gemini Workspace, since these services were aimed at commercial use. Nor does it include the commercial ToS for Claude, Grok, Le Chat, or Perplexity since these only applied to commercial use.
#JurisdictionCountServices
1.Customer's local courts7ChatGPT (basic), Claude (consumer), Copilot (basic), Gemini (basic), Grok (consumer), Le Chat (consumer), Meta AI
2.Arbitration2Ernie Bot, Perplexity (consumer)
3.China1DeepSeek
4.Singapore1Qwen Chat
5.Switzerland1Lumo
Details
  • Overview. All the ToS we surveyed asserted that any disputes arising under the contract must either be brought before the courts of a certain jurisdiction or be resolved through arbitration.
  • Common approach. The choice of forum typically matched the choice of law and the location of the provider's contracting party discussed above. Thus, under the ToS for Amazon Q, customers must sue Amazon before courts in Luxembourg;26Amazon Q CA, S.11(5), see Table "Governing Laws and Governing Courts". while Lumo customers must sue Proton in Switzerland.27Lumo ToS, S.10. Similarly, the ToS for DeepSeek and Qwen Chat specified courts in China and Singapore respectively.28DeepSeek ToU, S.9(2); Qwen Chat ToS, S.XII.
  • Grok. By contrast, while the ToS for Grok specified a contracting party in the State of Nevada and chose the law of the State of Tennessee for commercial use, they chose the courts in Tarrant County, Texas.29Grok ToS-E, S.14(6).
  • Consumer v. commercial use. In some cases, the choice of forum differed depending on whether the customer was a consumer or a business. Thus, under the Copilot (basic) ToS, consumers could choose to sue Microsoft before their local courts or before Irish courts.30Copilot MSA, S.10. Business customers of Microsoft 365 Copilot, however, had to sue Microsoft Ireland before Irish courts and Microsoft Inc before courts in Washington State.31Microsoft 365 Copilot MCA, under "Miscellaneous" (j). Similarly, under Le Chat ToS, consumers could sue Mistral before their local courts, while business customers had to sue Mistral in France.32Le Chat ToS (consumer), S.11; Le Chat ToS (commercial), 14(12)(3). Consumers could also sue Grok before their local courts.33See Grok ToS-C, under "Europe Specific Terms", "EST Venue of Jurisdiction". As with the choice of law above, the situation was more complicated in the case of Meta AI, since the choice of forum depended on both the customer's location and their status. The Meta AI ToS (consumer) stated that both UK and EU consumers could sue Meta before their local courts.34Meta AIs UK ToS, S.7; Meta AIs Europe ToS, S.7. The Meta AI ToS (commercial) stated that UK customers must sue Meta before courts in the State of California; while EU customers must sue Meta before Irish courts.35Meta AIs UK ToS, S.7; Meta AIs Europe ToS, S.7. The ToS of ChatGPT (basic) similarly chose local courts for consumers, but the courts in the State of California for commercial use.36OpenAI Europe ToU under "Court" and "Governing law (business use)".
  • Gemini. By contrast, the ToS for Gemini (basic) took a simpler approach. The ToS permitted all customers (i.e. both consumers and businesses) to sue Google before their local courts.37Gemini UK ToS and Gemini EU ToS, under "Settling disputes, governing law, and courts". These terms stood out from the others we surveyed, since they permitted business customers to sue the providers before their local courts. In other ToS, this approach was typically restricted to consumers.
  • Arbitration. Finally, five of the ToS we surveyed contained exclusive arbitration clauses. Thus, the ToS for Claude (commercial), Ernie Bot, Gemini Workspace, ChatGPT Enterprise, and Perplexity (consumer) stated that customers were bound to resolve disputes with the provider through arbitration only.38Chat-GPT Enterprise OpenAI SA, S.15; Claude Commercial ToS, S.J(2); Ernie Bot ToS, S.9(1); Gemini RM, S.I; Perplexity ToS-C, S.9. Three of these ToS were for services intended for enterprise or commercial use, namely: Claude (commercial), Gemini Workspace, and ChatGPT Enterprise. This was in contrast to the ToS for Claude (consumer), Gemini (basic), and ChatGPT (basic), which did not contain arbitration clauses.39Chat-GPT Europe ToU, under "Dispute Resolution"; Claude Consumer ToS, S.13; Gemini UK ToS and Gemini EU ToS, under "Settling disputes, governing law, and courts". By contrast, the Ernie Bot service was intended exclusively for consumer use.40Ernie Bot ToS, S.3(2). Further, Perplexity's arbitration clause applied only to consumer use; under the Perplexity (commercial) ToS, business customers could sue the provider before the courts of California.41Perplexity ToS-C, S.5(1), 9; Perplexity ToS-E, S.10(1). Note that the binding arbitration clause permitted the consumer to opt out of arbitration by email within 30 days under S.9(6).
  • Analysis. For an analysis of whether such choice of forum and arbitration clauses are binding, including under European consumer protection law, download our full paper free from SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698.
2. Provider Breach and Liability
2.1 General Duties of Care
Terms addressing a provider's general duty of care1The total count for the table is 20, since it covers all 20 ToS we surveyed, including both consumer and commercial ToS.
#ClauseCountServices
1.Disclaim any duty of care12Amazon Q, Claude (consumer), Claude (commercial), Microsoft 365 Copilot, DeepSeek, Ernie Bot, Grok (consumer), Gemini Workspace, Lumo, Perplexity (consumer), Perplexity (commercial), Qwen Chat
2.Duty to exercise reasonable care and skill or professional diligence4Copilot (basic), Gemini (basic), ChatGPT (basic), Meta AI
3.Duty to provide service that conforms to documentation4ChatGPT Enterprise, Grok (commercial), Le Chat (consumer), Le Chat (commercial)
Details
  • Overview. All of the ToS we surveyed addressed a provider's general duty of care under the contract. Broadly speaking, we observed three different approaches.
  • Complete disclaimers. First, twelve of the ToS we surveyed had terms which disclaimed completely any general duty of care. For example, the Claude ToS (consumer) and (commercial) stated that the service was provided "as is", "as available", and "without warranty of any kind", while expressly disclaiming "all implied warranties", including those arising "from statute".2Claude Commercial ToS, S.L(2); Claude Consumer ToS, S.11. The ToS for Amazon Q, Grok (consumer), DeepSeek, Ernie Bot, Lumo, Perplexity (consumer) and (commercial), and Qwen Chat all contained similar clauses, committing only to providing the service "as is", "with all faults", "without warranties or conditions of any kind", including regarding "satisfactory quality" or "fitness for a particular purpose".3Amazon Q CA, S.8; DeepSeek ToU, S.7(3); Ernie Bot ToS, S.8(1); Grok ToS-C, under "Disclaimer of Warranties"; Lumo ToS, S.7; Perplexity ToS-C, S.8(1); Perplexity ToS-E, S.7(1); Qwen Chat ToS, S.VII-VIII.
  • Examples: Copilot and Gemini. The ToS for Microsoft 365 Copilot and for Gemini Workspace similarly disclaimed any warranties, whether "express, implied, or statutory", including of quality or fitness for a particular purpose.4Copilot MCA, under "Warranties"(c); Gemini Workspace ToS, S.11. Microsoft added that the customer was "solely responsible" for "any application or AI agent it creates using or for use with Microsoft AI Services", including for compliance with "legal, regulatory, or licensing requirements".5Copilot ULT, under "Responsible Use of Microsoft AI Services". Finally, the terms specifically warned the customer that the service was not designed or intended for "high-risk use", that is: any use where "a service interruption" could result in "death or serious bodily injury" or "physical or environmental damage". Any such high-risk use by the customer was "at its own risk" and the customer indemnified Microsoft from any damages arising from such high-risk use.6Copilot ULT, under "High-Risk Use".
  • Conform to service documentation. Second, four ToS contained terms by which the provider committed to providing a service that conformed to the service documentation. The terms for ChatGPT Enterprise warranted that "when used in accordance with this Agreement", the service would "conform in all material respects with the documentation we provide to you or otherwise make publicly available".7Chat-GPT OpenAI SA, S.12(1). Beyond that, OpenAI disclaimed any further duties, in language similar to the clauses above.8Chat-GPT OpenAI SA, S.12(2). The ToS for Le Chat (consumer) and (commercial) and Grok (commercial) contained similar clauses.9Grok ToS-E, S.7(1) and the Documentation provided at https://docs.x.ai/docs/overview; Le Chat ToS (consumer), S.5; Le Chat ToS (commercial), S.7(2). In such cases, the provider's duties depended on the service description in the relevant documentation, which was in effect incorporated into the contract. The Le Chat ToS (consumer) added that EU consumers enjoyed "mandatory consumer protection regulations", which could not be overridden by the terms, and provided an appendix with specific terms for consumers in certain EU Member States (including France, Germany, and Italy).10Le Chat ToS (consumer), S. 5, 12, Appendix 2.
  • Reasonable care and skill. Third and finally, in four of the ToS we surveyed, providers committed to providing the service with reasonable care and skill or with professional diligence. In some cases, these terms applied to both business and consumer use. For example, the ToS for Gemini (basic) in the UK and for Meta AI contained a simple general duty of care, namely to provide the service "using reasonable skill and care"11Gemini UK ToS, under "Warranty". The ToS for European customers differed. See e.g. the Gemini EU ToS under "Legal guarantee". or to "exercise professional diligence".12MetaAI UK ToS, S.6(B); MetaAI Europe ToS, S.6(B). In other cases, the terms imposed a duty of care which applied only to consumer use of the service. The ToS for Copilot (basic) recognised that Microsoft had "an obligation [...] to provide the Services using reasonable care and skill". Beyond that, Microsoft committed only to providing the service "as is", "as available", and "with all faults", while excluding any further implied warranties.13Copilot MSA, S.11. The clause contained an exception for defects "hidden [...] in bad faith" or which "rendered use of the services impossible". The terms that governed ChatGPT (basic) also imposed a duty on OpenAI to provide the service "with reasonable skill and care" and "with professional diligence".14Chat-GPT Europe ToU, under "Our commitments to you".
2.2 Accuracy of Output
  • Overview. A common concern for customers using GenAI services is that the outputs might be inaccurate. Large language models generate outputs based on probabilities. As a result, some outputs will be probable (based on training data and other inputs), but inaccurate. Such outputs are also referred to as "hallucinations".15See e.g. S. Wachter, B. Mittelstadt, and C. Russell, "Do large language models have a legal duty to tell the truth?" R. Soc. Open Sci. (2024), pp.2, 5, 9. All of the ToS we surveyed contained a disclaimer regarding the accuracy of service outputs. All disclaimed any contractual commitment that their service would generate accurate or error-free outputs.16Amazon Q CA, S.8; Chat-GPT ToU, under "Accuracy"; Chat-GPT SA, S.12(2); Claude Commercial ToS, S.L(2); Claude Consumer ToS, S.4; Copilot MSA, S.11; DeepSeek ToU, S.7(3); Ernie Bot ToS, S.4(2); Gemini UK ToS, under "Disclaimers"; Gemini Workspace SST, S.12(1); Grok ToS-C, under "User Content"; Grok ToS-E, S.3(1); Lumo ToS, S.3; MetaAI UK ToS, S.6(A); MetaAI Europe ToS, S.6(A); Le Chat ToS (consumer), S.4; Le Chat ToS (commercial), S.3(5); Perplexity ToS-C, S.8(1); Qwen Chat ToS, S.VII(d).
  • Detailed warnings. Some ToS went further, providing detailed warnings to customers. For example, the Claude ToS (consumer) warned consumers that large language models were "frontier technologies" that were "still improving in accuracy, reliability and safety". Their outputs might contain "material inaccuracies even if they appear accurate because of their level of detail or specificity". Consumers should not rely on outputs "without independently confirming their accuracy".17Claude Consumer ToS, S.4. The Claude (commercial) terms similarly warned business customers that outputs might be "false, incomplete, misleading or not reflective of recent events", and so "factual assertions in Outputs should not be relied upon without independently checking their accuracy".18Claude Commercial ToS, S.D(3). Meta AI's ToS similarly stated that GenAI was "a new technology" which was "still improving" and made no guarantees that the service would be "safe, secure or error-free". Meta was "not responsible for Outputs"; instead the customer was "responsible for checking Outputs for accuracy and suitability".19MetaAI UK ToS, S.6(A); MetaAI Europe ToS, S.6(A). The ToS for Amazon Q contained similar warnings.20Amazon Q RP, under "Responsible AI Requirements". In addition, the ToS for Amazon Q stated that the service was not intended for use in "hazardous environments" or "critical systems", adding that the customer was "responsible for liability that may arise in connection with any such uses".21Amazon Q AWS ToS, S.50(6).
  • Examples: Copilot and Gemini. The ToS for Copilot (basic) stated that the service "can make mistakes" and was "for entertainment purposes only". The customer should not "rely on Copilot for important advice". Further, Microsoft did not "promise that any Copilot's Responses won't infringe someone else's rights (like their copyrights, trademarks, or rights of privacy) or defame them". The customer was "solely responsible" if they chose to "publish or share Copilot's Responses publicly or with any other person".22Copilot ToU, under "Important disclosures & warnings". The UK ToS for Gemini (basic) and for Gemini Workspace similarly warned the customer not to rely on the service as a substitute "for medical, legal, financial, or other professional advice";23Gemini UK ToS, under "Disclaimers"; Gemini Workspace SST, S.12(1). while the Perplexity ToS (consumer) and (commercial) warned customers more broadly not to rely on the service for "advice of any kind".24Perplexity ToS-C, S.8(1)(a); Perplexity ToS-E, S.7(2).
  • Examples: Qwen Chat, Ernie Bot, Grok. Further, the ToS for Qwen Chat, Ernie Bot, and for Grok (consumer) all stated that use of outputs was entirely at the customer's risk and warned the customer not to rely on outputs as the "sole source of truth or factual information" or as "professional advice".25Grok ToS-C, under "Disclaimer of Warranties"; Ernie Bot ToS, S.4(2); Qwen Chat ToS, S.VII(d). The Qwen Chat ToS added that since outputs might be "inaccurate, misleading, incomplete, erroneous, or lack context" and contain "inappropriate, unintended, and/or offensive elements", the customer was "solely responsible" for "reviewing the Outputs" and verifying their "accuracy" and "fitness for [a] specific use case".26Qwen Chat ToS, S.III(4).
  • Tension with promotional materials. These clauses were often at odds with a provider's public statements advertising a service, including on its website.27See also S. Wachter, B. Mittelstadt, and C. Russell, "Do large language models have a legal duty to tell the truth?" R. Soc. Open Sci. (2024), pp.3-4. For example, on the Copilot webpage, Microsoft promoted the model's ability to "conduct research";28See https://www.microsoft.com/en/microsoft-copilot/for-individuals [visited on 3 February 2026]. on the Claude webpage, Anthropic promised "expert-level collaboration [...] from coding a product to critical data analysis";29See https://claude.com/product/overview [visited on 3 February 2026]. while the webpage for Perplexity Enterprise promised "[r]eal research, real answers, in real time [...] for confident decisions".30See https://www.perplexity.ai/enterprise [visited on 22 January 2026]. Similarly, on the Qwen Chat webpage, Alibaba offered "advanced reasoning and problem-solving [...] with clarity and precision, leveraging real-time internet data to deliver comprehensive, logical, and actionable solutions".31See https://qwen.ai/home [visited on 3 February 2026]. It was not clear whether warnings in the ToS were effective in tempering user expectations. Survey evidence suggested that some two-thirds of respondents trusted the technical ability of AI systems to provide accurate and reliable output.32KPMG, "Trust, attitudes and use of artificial intelligence" (2025), p. 27 https://assets.kpmg.com/content/dam/kpmgsites/xx/pdf/2025/05/trust-attitudes-and-use-of-ai-global-report.pdf. The percentage of respondents trusting AI differed per country, see p.28. Further, two-thirds of employees reported having relied on AI output at work without critically evaluating the information it provided.33Ibid., p.75. This implied that many users relied on the accuracy of GenAI outputs, despite providers' stark warnings of inaccuracy in the ToS.
2.3 Restrictions and Exclusions of Liability
Provider liability terms for commercial use34The total count for the table is 13. It does not include Alibaba Qwen Chat or Ernie Bot, since the services were for non-commercial use only. Nor does it include the consumer ToS for Claude, Copilot (basic), Grok, Le Chat, or Perplexity, since these applied only to consumer use.
#ClauseCountServices
1.Exclude indirect losses incl. lost profits4Claude (commercial), Le Chat (commercial), ChatGPT (basic), Grok (commercial)
2.Exclude indirect losses incl. lost profits, except for gross negligence / wilful misconduct3ChatGPT Enterprise, DeepSeek, Perplexity (commercial)
3.Exclude indirect losses + lost profits2Amazon Q, Gemini Workspace
4.Exclude indirect losses + lost profits, except for gross negligence / wilful misconduct2Microsoft 365 Copilot, Gemini (basic)
5.Exclude all liability1Lumo
6.No exclusion for breach of contract or negligence1Meta AI
Provider liability terms for consumer use35The total count for the table is 12. It does not include Amazon Q, Chat-GPT Enterprise, Microsoft 365 Copilot, or Gemini Workspace, since these services were aimed at commercial use. Nor does it include the commercial ToS for Claude, Grok, Le Chat, or Perplexity since these only applied to commercial use.
#ClauseCountServices
1.No exclusion in case of breach of contract or negligence6ChatGPT (basic), Claude (consumer), Gemini (basic), Grok (consumer), Le Chat (consumer), Meta AI
2.Exclude all liability4Qwen Chat, Ernie Bot, Lumo, Perplexity (consumer)
3.Exclude indirect losses, except for (gross) negligence or intentional conduct2DeepSeek, Copilot (basic)
Details
  • Overview. All the ToS we surveyed had terms addressing the provider's liability for damages. Liability clauses can be complicated and require close reading to understand which types of losses they exclude. For details on our approach to interpreting liability clauses, please see our full paper at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698. Broadly speaking, we observed four different approaches to provider liability: (1) excluding liability to pay damages entirely; (2) excluding liability to pay damages for indirect losses only (while being silent as to liability for direct losses); (3) excluding liability to pay damages for indirect losses, other than for losses stemming from the provider's gross negligence or wilful misconduct; and, (4) stating that the provider was liable (or at least explicitly not excluding such liability) for damages in case of breach of contract, (mere) negligence, or failure to exercise reasonable care and skill.
  • Exclude all damages. In the first category, four of the ToS we surveyed excluded all damages in the broadest possible terms, including both direct and indirect losses. The Qwen Chat ToS excluded both "direct" and "indirect" damages, as well as "any damages whatsoever", resulting from "any matters relating to services".36Qwen Chat ToS, S.VIII(1). The Ernie Bot ToS contained a similar clause.37Ernie Bot ToS, S.8(2). The ToS for Lumo and Perplexity (consumer) similarly excluded liability for "any damages whatsoever" related to the use of the service.38Lumo ToS, S.7; Perplexity ToS-C, S.8(2). These sweeping terms appeared to exclude any provider liability to pay damages for losses caused by a breach of contract.
  • Exclude indirect losses. Second, five of the ToS we surveyed contained terms excluding all indirect losses, without explicitly mentioning direct losses. For example, the ToS for Claude (commercial) excluded liability for indirect losses, including lost profits (etc.), but were silent as to direct losses.39Claude Commercial ToS, S.L(3). The ToS that governed Le Chat (commercial), and Grok (commercial) contained similar clauses, as did the ToS for commercial use of ChatGPT (basic).40Le Chat ToS (commercial), S.9(1); Grok ToS-E, S.8(1); Chat-GPT Europe ToS, under "Business use of the Services addendum". These clauses left open the possibility that providers might be liable to pay damages for direct losses (including lost profits), insofar as such liability arises under the governing law. Conversely, the terms clearly excluded liability for indirect damages stemming from a breach of contract. In this respect, the Grok ToS (commercial) added that, in case the provider breached its duty to provide the service in accordance with the service documentation, the customer's "sole and exclusive remedy" was for xAI to "use commercially reasonable efforts" to correct the non-conformity "in xAI's sole discretion", or to allow the customer to terminate their use of the service.41Grok ToS-E, S.7(1).
  • Amazon Q and Gemini. The ToS for Amazon Q took a slightly different approach. As above, the terms excluded liability for "indirect" or "consequential" losses, without mentioning direct losses. Yet the terms separately excluded "loss of profits, revenues, customers, opportunities, or goodwill" and liability for "the value of [customer] content" -- not as an example of indirect losses, but as a type of harm in itself.42Amazon Q CA, S.9(1). The relevant clause excluded "liability [...] under any cause of action or theory of liability, even if a party has been advised of the possibility of such liability, for (a) indirect, incidental, special, consequential or exemplary damages, (b) the value of your content, (c) loss of profits, revenues, customers, opportunities, or goodwill, or (d) unavailability of the services or AWS content". The ToS for Gemini Workspace similarly excluded liability for "indirect" or "consequential losses" and separately for "lost revenues, profits, savings or goodwill".43Gemini Workspace ToS, S.12(1). This wording implied that the provider excluded liability for lost profits and customer content in all cases (i.e. whether the loss was direct or indirect). In this respect, Amazon Q and Gemini Workspace's exclusion clauses appeared to be broader than the terms of Claude (commercial) and the other providers we reviewed above.
  • Exclude indirect losses, except gross negligence. Third, five of the ToS we surveyed excluded liability for indirect losses only, but contained an exception for such losses when caused by the provider's conduct in cases of gross negligence or wilful misconduct. Such clauses left open the possibility that a provider would be liable in case of (1) direct losses, or (2) indirect losses caused by gross negligence, while excluding liability in other cases (such as indirect losses caused by 'mere' or 'simple' negligence). For example, the terms that governed ChatGPT Enterprise excluded "indirect" or "consequential [...] damages (including lost profits)", except in cases of "gross negligence or willful misconduct".44Chat-GPT Enterprise SA, S.14(1). The Copilot (basic) ToS similarly stated that Microsoft would not be liable "for any indirect damage, including financial loss, such as loss of profit", unless it had "acted with gross negligence or willful misconduct".45Copilot MSA, S.12(B)-(C).
  • DeepSeek. The ToS for DeepSeek took a variation on this approach. As above, the terms did not mention direct losses and excluded liability for indirect losses (including loss of profits or goodwill). However, the terms contained an exception for losses caused by "intentional or negligent conduct".46DeepSeek, ToU S.7(4). This term implied a lower standard for provider liability, namely: it left open the possibility that the provider would be liable for indirect losses caused by 'mere' negligence (as opposed to 'gross' negligence). The term applied equally to commercial and consumer use (while recognising that consumers could not be deprived of their rights under consumer protection law).47DeepSeek ToU, S.7(1). By contrast, the Perplexity ToS (commercial) excluded liability for indirect losses (including lost profits, etc.), except in the case of wilful misconduct -- but made no mention of negligence, whether gross or otherwise.48Perplexity ToS-E, S.7(3).
  • Copilot. The ToS for Microsoft 365 Copilot took a slightly different approach. The exclusion clause explicitly limited Microsoft's liability to "direct damages" only, while excluding liability for "indirect" damages. The ToS separately excluded liability for "loss of [...] profits [...] (whether direct or indirect); or [...] loss of business information". In both cases, the limitation did not apply in case of gross negligence or wilful misconduct.49Copilot MCA, under "Limitation of liability", S.(e) and (f). The clauses read: "In no event will either party be liable for indirect, incidental, special, punitive, or consequential damages; loss of revenue, profits, or anticipated savings (whether direct or indirect); or loss of use, loss of business information, or interruption of business, however caused or on any theory of liability" and "No limitation or exclusions under this Agreement will apply to liability arising out of either party's [...] willful misconduct or gross negligence". As with Amazon Q's ToS above, the clauses excluded lost profits not as an example of indirect losses, but as a type of harm in itself (i.e. whether direct or indirect).
  • No exclusion for foreseeable losses stemming from provider negligence. Fourth, seven of the ToS we surveyed contained clauses which either explicitly stated that the provider was liable for foreseeable losses stemming from the provider's breach of contract, negligence, or failure to exercise reasonable care and skill; or did not attempt to exclude such liability. Providers typically limited this approach to consumers, while excluding damages more broadly in case of commercial use (as summarised above). For example, the Claude (consumer) ToS stated that the provider was liable for foreseeable losses resulting from its breach of contract or its negligence.50Claude Consumer ToS, S.11. Per the terms, consumers further agreed not to use the service for "commercial or business purposes" and Anthropic accordingly excluded "any loss of profit, loss of business, business interruption, or loss of business opportunity".51Claude Consumer ToS, S.11. The Le Chat ToS (consumer) took a similar approach, stating that Mistral's liability was limited to foreseeable losses resulting from a failure to exercise reasonable care and skill or breach of contract.52Le Chat ToS (consumer), S.4. The ToS for consumer use of ChatGPT (basic) and for Grok (consumer) took a similar approach. The terms explicitly did not exclude liability for reasonably foreseeable losses caused by breaches of contract and where the provider had failed to act with professional diligence.53The two clauses were nearly identical, stating that: "provided that we have acted with professional diligence, we do not take responsibility for loss or damage caused by us, unless it is caused by our breach of these Terms or [is] reasonably foreseeable at the time of entering into these Terms". See Chat-GPT Europe ToS, under "Our commitments to you"; Grok ToS-C, under "Regional Specific Terms", "EST Limitation of Liability".
  • Gemini and MetaAI. The UK ToS for consumer use of Gemini (basic) appeared to take a similar approach, albeit less explicitly. The ToS stated that Google was "only liable for its breaches of these terms [...] subject to applicable law".54Gemini UK ToS, under "Liabilities". The approach for European businesses differed, as Google had different ToS documents for customers in different countries. See e.g. Gemini ToS for German users at https://policies.google.com/terms?hl=en-GB&gl=DE; or French users at https://policies.google.com/terms?hl=en-GB&gl=FR. This clause implied that UK consumers would be entitled to damages for breaches of contract in line with the general rules on damages under English contract and consumer protection law. Finally, the ToS for Meta's MetaAI services took a single approach that applied to both consumer and commercial use. The MetaAI ToS implicitly did not exclude liability for foreseeable losses caused by the provider's breach of contract and where the provider had failed to act with professional diligence.55Meta UK ToS, S.6(b); Meta Europe ToS, S.6(b). The relevant clause read: "Provided that we have acted with [...] professional diligence, we do not accept responsibility for losses not caused by our breach of these Terms or otherwise by our acts, losses that are not reasonably foreseeable by you and us at the time of entering into these Terms, and events beyond our reasonable control".
  • Analysis. For an analysis of the clauses disclaiming duties and restricting or excluding liability in both business-to-business and business-to-consumer contexts under UK law (including the Unfair Contract Terms Act 1977 and the Consumer Rights Act) and European law (including the EU Data Act; the Product Liability Directive; the Digital Content Directive; and the Unfair Contract Terms Directive), download our full paper free from SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698.
2.4 Liability Caps
  • Overview. The vast majority of the ToS we surveyed contained terms that capped the provider's liability, especially for business customers. The cap was typically set based on the fees paid by the customer over a certain time period or at an absolute amount. Thus, the ToS for Amazon Q, Microsoft 365 Copilot, and Gemini Workspace capped liability at the fees paid by the customer during the preceding twelve months -- or at USD $5,000, whichever was greater, in the case of the latter two services.56Amazon Q CA, S.9(2); Copilot MCA, under "Limitation of liability" (b); Gemini Workspace, S.12(2). The Qwen Chat ToS similarly capped liability at twelve months' fees, or EUR €100 / USD $100 (whichever was greater);57Le Chat ToS, S.6; Qwen Chat ToS, S.VIII(2). while the Ernie Bot ToS capped liability at six months' fees or USD $100 (whichever was greater).58Ernie Bot ToS, S.8(2). By contrast, the Lumo ToS capped provider liability at USD $100 or "the amount you paid us, if any, for use of your Account or the Services, whichever amount is greater" -- without specifying a time frame for the fees.59Lumo ToS, S.7. In any event, all such terms had the effect of tying the provider's liability to the customer's spend: higher-spend customers were entitled to higher damages. Conversely, customers using a free service faced a low cap of only USD $100.
  • Consumer v. commercial use. In some cases, the cap differed between businesses and consumers. For example, the ToS for Grok (commercial), Le Chat (commercial), and MetaAI (commercial) capped provider liability at the fees paid by the customer during the preceding 12 months;60Grok ToS-E, S.8(2); Le Chat ToS (commercial), S.9(2); MetaAI Commercial Terms, S.4. For services provided free of charge, the MetAI ToS set a cap of USD $100. while the ToS which applied to European consumers of the same services did not cap provider liability at all.61See Grok ToS-C, under "Limitation of Liability" and under "EST Limitation of Liability"; Le Chat ToS (consumer); MetaAI ToS. The Gemini (basic) ToS similarly capped liability only for business and organizational customers (at 125% of the fees paid by the customer during the preceding 12 months or GBP £500, whichever was greater); not for consumers.62Gemini UK ToS under "For business users and organisations only"; Gemini EU ToS under "For business users and organisations only". Similarly, the Copilot (basic) ToS did not contain a cap on provider liability in case of consumer use.63Copilot MSA; Copilot ToU. By contrast, the ToS for Claude (consumer) and Perplexity (consumer) provided a lower cap for consumer use (six months' fees or GBP £100, whichever was greater) than the ToS for Claude (commercial) and Perplexity (commercial) provided for commercial use (12 months' fees).64Claude Commercial ToS, S.L(3)(a); Claude Consumer ToS, S.11; Perplexity ToS-C, S.8(2); Perplexity ToS-E, S.7(3).
  • ChatGPT Enterprise. Alone among the ToS we surveyed, OpenAI's ChatGPT Enterprise service did not cap provider liability for commercial use in case of gross negligence or wilful misconduct. Instead, the terms capped provider liability at the usual rate of 12 months' fees, except in cases of provider gross negligence or wilful misconduct. In the latter case, no cap applied.65Chat-GPT Enterprise OpenAI SA, S.14(2). This differed from the terms that governed its ChatGPT (basic), which did not cap liability for consumer use and capped liability for commercial use in all cases at the usual rate of 12 months' fees or USD $100 -- whichever was greater (i.e. without an exception for gross negligence or wilful misconduct).66Chat-GPT Europe ToU, under "Our commitments to you" and "Business use of the Services addendum".
  • DeepSeek. Lastly, alone among the ToS we surveyed, the DeepSeek ToS did not contain a cap on liability for commercial use at all.67See DeepSeek ToU, S.7. Further, as described above, the ToS did not appear to exclude liability for direct losses, nor for indirect losses stemming from intentional or negligent conduct. Taken together, these terms suggest that the provider might be exposed to uncapped liability for direct losses, as well as indirect losses caused by negligent breaches of contract. That said, as discussed above, the DeepSeek ToS did contain a broad disclaimer of any duty of care. This might make it difficult for a customer to prove that the provider has in fact breached its duties under the contract in case the service malfunctions. In any event, in practice, the terms need to be interpreted under Chinese law, since the ToS chose Chinese law to govern disputes under the contract -- subject to any mandatory consumer protections that may arise under UK or EU law in the case of consumer use.
3. Customer Breach and Acceptable Use Policies
3.1 General Prohibitions
  • Overview. When surveying the general prohibitions in the ToS, we noted that the specific wording of the ToS differed. Some providers imposed brief, broadly worded, categorical restrictions,1Amazon Q AUP; DeepSeek ToS; Lumo ToS; Qwen Chat ToS. while others outlined intricate scenarios.2Claude UP; Copilot (basic) MSA, S.3; Ernie Bot ToS, S.3(2); Gemini Gen AI PUP; Le Chat UP; Perplexity AUP. Nonetheless, in substance, the prohibitions covered the same fundamental types of user conduct. For example, all 20 of the ToS we surveyed expressly banned users from using the service for any illegal activity, violence, terrorism, hate speech, incitement, harassment, fraud and circumvention of the security measures of a computer system (including the security of the GenAI service itself).3See, Amazon Q UP; ChatGPT UP; Claude UP; Copilot (basic) MSA, S.3; Ernie Bot ToS, S.3(2); Gemini GenAI PUP, S.1; Grok AUP; DeepSeek ToS, S.3(4); Le Chat UP; Lumo ToS, S.2; Meta AI Europe ToS, S.1(B); Perplexity AUP; Qwen Chat ToS, S.3(1). Additionally all ToS, with the exception of DeepSeek and Lumo, specifically prohibited spam, phishing, and privacy harms (e.g. doxing, unlawful tracking, impersonation).4Amazon Q UP; ChatGPT UP; Claude UP; Copilot (basic) MSA, S.3; Ernie Bot ToS, S.3(2); Gemini GenAI PUP, S.1; Grok AUP; Le Chat UP; Meta AI Europe ToS, S.1(B); Perplexity AUP; Qwen Chat ToS, S.II(1) and III(1). Instead, the DeepSeek ToS addressed these only impliedly, through its prohibitions on the violation of applicable laws, on infringing the legal rights of others, and on generating output designed to “antagonize”, “harass”, “scare” or “distress” others.5DeepSeek ToS, S.3(2) and (4). Similarly, the Lumo ToS could be construed as banning privacy harms only when the acts constituted “illegal or prohibited activities” or “harassing” or “abusing” others based on protected characteristics such as gender or sexual orientation.6Lumo ToS, (S.2).
  • System security. While all thirteen providers prohibited the circumvention of any computer system security features (including safety filters), the ToS for Claude, Ernie Bot, Copilot, and Qwen Chat went further by explicitly banning prompt injection7Qwen Chat ToS, S.III(1); Claude UP, under “Do Not Abuse our Platform”; Perplexity AUP, S.3(b). or prompt-based manipulations.8Copilot ToU, under “code of conduct”; Ernie Bot ToS, S.3(2)(i). The ToS for Amazon Q and Perplexity used broader language to prohibit any circumvention of “safety filters and functionality or prompt models”9Amazon Q RP under “Prohibitions”. or other attempts “to discover the source code, model weights or other algorithmic parameters of the Services”10Perplexity AUP, S.3(b). which in practice would cover attempts to “jailbreak” the model or bypass moderation systems. These additions may indicate a more mature recognition of the security risks which are specific to GenAI. In addition, some ToS addressed a specific risk relating to misinformation and authenticity, which is a widely acknowledged systemic risk associated with generative AI11See World Economic Forum, Global Risks Report 2024, “Misinformation and disinformation (#1) is a new leader of the top 10 rankings this year”. and a focal point of emerging regulatory frameworks that emphasise provenance and content credentialing mechanisms.12See Art. 50(2) EU AI Act, which requires providers of AI generated content to ensure that any output is clearly marked “in a machine-readable format and detectable as artificially generated or manipulated”. Another example is China where the Cyberspace Administration, through its AI Labeling Rules, mandates that any AI-generated content provider must clearly mark AI-generated content. The ToS for Copilot, Ernie Bot, and Perplexity were distinctive in expressly addressing this risk by prohibiting the tampering of content credentials and other provenance methods (“AI Content Credentials”),13Copilot CoC-E, under “Usage restrictions”; Ernie Bot ToS, S.3(2)(i). or simply watermarks.14Perplexity AUP, S.3(d).
  • Weapons manufacturing. While several ToS prohibited the use of their services for weapons manufacturing,15See e.g. Chat-GPT UP; Claude UP, under “Do Not Develop or Design Weapons”; Grok AUP, S.2(3); Le Chat UP, under “prohibited content and activities”. the ToS for Amazon Q specifically prohibited users from using its service to “perform a lethal function in a weapon without human authorisation or control”.16Amazon Q RP. Instead of addressing weapons development generally, this clause concentrated on outlawing autonomous deadly force. This left a grey area: the ToS did not specifically prohibit using Amazon Q in the creation of weapons generally — or to carry out deadly tasks specifically, provided there is human supervision.
3.2 High-risk or Regulated Use
  • Overview. An area of divergence in AUPs concerned the treatment of regulated or so-called high-risk practices. Here, the ToS we surveyed took two different approaches: either (i) an outright prohibition or (ii) a conditional permission subject to safeguards.
  • Outright prohibition. Five providers imposed outright prohibitions in their ToS. The ToS for Qwen Chat and Meta AI explicitly banned users from generating outputs that could amount to professional advice (e.g. medicine, finance, law, or pharmaceuticals),17Meta AI ToS S.2(c); Qwen Chat ToS, S.II-III. or that could be used in safety-critical contexts (e.g. life support, emergency services, nuclear facilities, autonomous vehicles, or traffic control systems).18Qwen Chat ToS, S.II-III. These restrictions precluded customers from relying on GenAI services in relation to activities that could generate significant harm. The ToS for Perplexity expressly provided that users should “avoid high risk activity”. It also went further than all other providers by introducing a specific ban on “any activity or practice that is prohibited or considered ‘high risk’ under the European Union’s AI Act”.19Perplexity AUP, S.2(f). The DeepSeek ToS provided that any use which “may have serious harmful impacts on physical health, psychology, society, or the economy” or any “regulated uses, including [...] dangerous environments or [...] critical systems” was not permitted under its terms.20DeepSeek ToS, S.3(6), 4(f). Finally, while the ToS for Lumo did not have a restriction regarding high-risk activities, it prohibited customers from using outputs that could have “a legal or material impact on that person, such as making credit, educational, employment, housing, insurance, legal, medical, or other important decisions about them”.21Lumo ToS, S.3.
  • Conditional permission. By contrast, five other providers permitted the use of the services in regulated or high-risk situations, as long as customers took specific precautions.22Chat-GPT UP, under “Protect people”; Claude UP, under “High Risk Use Case Requirements”; Ernie ToS, S.3(2)(vii); Le Chat UP, under “Prohibited content and activities”; Perplexity AUP, S.2(d). For example, the ToS for Claude identified certain high-risk use cases (e.g. legal, healthcare, insurance, finance, employment and housing, academic testing, accreditation and admissions, media or professional journalistic content) and required users to: (1) ensure a qualified professional was in-the-loop to review outputs before dissemination, and (2) to disclose to individuals or consumers directly that outputs were generated by AI.23Claude UP, under “High Risk Use Case Requirements”. The ToS for Ernie Bot provided for the same two-prong requirement, while adding that the “potential limitations” of the use of AI must also be disclosed by customers.24Ernie ToS, S.3(2)(vii). The ToS for ChatGPT took a similar approach. First, the terms prohibited the use of the service to generate “advice that requires a license, such as legal or medical advice”, when done “without appropriate involvement by a licensed professional”.25Chat-GPT UP, under “Protect people”. This left it to the customer to determine, at least in first instance, what level of involvement of a licenced professional would qualify as “appropriate”. Second, the ToS also restricted the use of the service for the “automation of high-stakes decisions in sensitive areas without human review”, and listed a number of such areas, including education, housing, employment, finance, insurance, legal, medical, national security, migration, and law enforcement.26Chat-GPT UP, under “Empower people”.
  • Grok. The ToS for Grok departed from the above by combining both approaches: an outright ban and a conditional permission. On the one hand, users must not make “high-stakes automated decisions that affect a person’s safety, legal or material rights, or well-being (such as making financial credit, educational, employment, housing, insurance, legal, medical, or other important decisions about or for them)”.27Grok ToS-C, under “What you cannot do”. On the other hand, the ToS prohibited the use of the service when “operating in a regulated industry [...] without complying with the relevant regulations”.28Grok AUP, S.1(5). This suggests that users may employ Grok to provide services within a regulated profession, such as legal services, only if they themselves meet the applicable regulatory requirements, for example by being a duly licensed attorney. We also found this peculiarity in Perplexity’s ToS. As explained above, the ToS for Perplexity expressly warned against high-risk activity, i.e., “to conduct or facilitate any activity that poses a significant threat to individuals or society”.29Perplexity AUP, S.2. Nevertheless the ToS also provided that customers must not provide “medical, financial, tax, legal or other services that would typically require specialized credentials or licensing without appropriate professional oversight”.30Perplexity AUP, S.2(d).
  • Amazon Q, Gemini, and Copilot. The ToS for Amazon Q, Gemini, and Copilot adopted a more general conditional permission regarding the use of their services to make “decisions or take actions without appropriate human oversight in applications that could have a consequential impact on an individual’s legal or financial position, life or employment opportunities, human rights, or that may result in physical or psychological harm” (or similar wording to that effect).31Amazon Q RP, under “Responsible AI Requirements”; Copilot CoC-E, under “Usage Restrictions”, S.5; Gemini GenAI PUP, S.1(h). This approach reflects that taken in the GDPR, which grants individuals the right not to be subjected to “a decision based solely on automated processing, [...] which produces legal effects concerning him or her or similarly significantly affects him or her”.32Art 22(1) GDPR.
3.3 Exceptions
  • Overview. While all 13 of the providers we surveyed maintained broad prohibitions on harmful or unlawful use, only five providers carved out narrow exceptions to these prohibitions in their ToS. These exceptions appeared to apply only in limited contexts. The ToS for Gemini and Perplexity stated that the provider may make exceptions to its AUP based on “educational, documentary, scientific, or artistic considerations”, or where “harms are outweighed by substantial benefits to the public”.33Gemini Gen AI PUP; Perplexity AUP, S.5. The ToS for Copilot further allowed enterprise customers to generate otherwise prohibited content, if the purpose was to evaluate, train, or fine-tune safety systems for their own use; to test Microsoft’s AI services under the company’s penetration testing rules; or to fine-tune models to block harmful content in future use. Customers could use any harmful output created under these exceptions internally only, for evaluation or reporting, and must never share it externally.34Copilot CoC-E, under “Limited Exception”. The ToS for xAI’s Grok service allowed official red teams to test the security of its GenAI system, which would otherwise violate the general prohibitions on circumventing security measures.35Grok AUP, S.3.
  • Claude. The ToS for Anthropic’s Claude service did not provide a directly applicable exception. Instead, the ToS stated that Anthropic might enter into separate contracts which permitted limited tailoring of its use restrictions for certain public-sector customers, if the provider deemed the contractual safeguards and risk mitigation measures adequate.36Claude UP. Thus, public-sector customers could contact Anthropic to negotiate such separate contracts. By contrast, the other eight providers appeared not to provide specific exceptions to their general prohibitions.37These services were Amazon Q, ChatGPT, DeepSeek, Le Chat, Lumo, Meta AI, and Qwen Chat.
3.4 Disclosure Obligations
  • Overview. Beyond prohibitions, five of the thirteen providers we surveyed included in their AUPs express provisions that required users to take affirmative actions. The most common was a general disclosure requirement, i.e. the customer must make it clear that output was AI-generated in certain circumstances, beyond the high-risk use cases discussed above. This obligation entailed both: (i) not misrepresenting AI-generated output as having been created by humans; and, (ii) requiring disclosure in public-facing or consumer-facing settings where individuals interact with GenAI systems (e.g. chatbots).38Claude UP, under “High-Risk Use Case Requirements”; Copilot CoC-E, under “Responsible AI requirements”; DeepSeek ToU, S.3(1); Perplexity AUP, S.2(b). This was likely driven by regulatory and risk management considerations. The EU AI Act, for example, explicitly requires such transparency whenever individuals interact with an AI system.39EU AI Act, Art. 50. Such terms might also reflect a business decision: GenAI providers might require their customers to provide transparency to end-users, as a way to maintain end-user trust in AI systems generally and to reduce risks from fraud or misrepresentation.
  • Perplexity. Perplexity was the only provider in our sample whose ToS imposed a disclosure obligation which required customers to inform end-users when they were interacting with an AI chatbot, unless “it is obvious from the context”.40Perplexity ToS, S.2(b). This required customers to determine when an average user could reasonably be expected to recognize that they are communicating with an AI system. Such an assessment becomes increasingly difficult as AI technologies advance and the boundary between human and machine-generated interactions becomes less perceptible.
  • No express obligation. By contrast, there were no express disclosure obligations in the ToS of the remaining providers.41The services were Amazon Q, Chat-GPT; Gemini, Grok, MetaAI, Le Chat, Lumo and Qwen Chat. However, the ToS for Ernie, Gemini, Grok, and Le Chat adopted a negative framing of the same principle: rather than requiring users to affirmatively disclose AI involvement, they prohibited users from misleading others about the provenance of AI-generated content, such as by suggesting it was produced by a human.42Ernie Bot ToS, S.3(2)(vi); Gemini ToS, under “Don’t abuse our services”; Grok AUP, S.3; Le Chat ToS, S.3.
3.5 Misinformation / Disinformation
  • Overview. Ten of the thirteen providers we surveyed specifically prohibited the use of the services to promote, engage in, or spread so-called ‘misinformation’ or ‘disinformation’.43Amazon Q RP, under “Prohibitions”; Claude UP, under “Do Not Create or Spread Misinformation”; CoPilot MSA, under “AI Services”; DeepSeek ToS, S.3(1); Ernie Bot ToS, S.3.(2)(vi); Le Chat UP, under “Usage Policy”, “Misinformation”; Gemini GenAI PUP, S.4(c)-(d); Meta AI UK ToS, S.1(B); Perplexity AUP, S.2(a); Qwen Chat ToS, S.2(8). When defining such mis- or disinformation, the terms typically referred to creating output with the intent to deceive others,44See e.g. Amazon Q RP, under “Prohibitions”; CoPilot (basic) MSA, under “AI Services”, S.(ix)(2); Le Chat UP, under “Misinformation”; Gemini GenAI PUP, S.4(c)-(d). and in some cases referred specifically to disseminating “conspiratorial narratives” or “conspiracy theories”.45Claude UP, under “Do Not Create or Spread Misinformation”; Le Chat UP, under “Misinformation”.
  • ChatGPT and Grok. By contrast, the ToS for ChatGPT and Grok did not explicitly call out misinformation as a prohibited practice. That said, the ChatGPT ToS did prohibit using the service to “manipulate or deceive people” generally, which could cover some of the same conduct.46Chat-GPT UP. The ToS for Grok only prohibited using the service for the purpose of “[m]isleading others or not being transparent regarding your use of AI” insofar as doing so breached applicable laws or regulations.47Grok ToS-C, under “Using our Service”. See also Grok AUP under “Respect guardrails and don’t mislead”. This might have reflected owner Elon Musk’s strong views on freedom of speech.48See e.g. M. Pollet, “Elon Musk’s X tells the EU: We’re a safe space for free speech”, 28 November 2024, https://www.politico.eu/article/elon-musks-x-eu-safe-space-free-speech-digital-services-act/. Lastly, the ToS for Lumo did not address mis- or disinformation at all.49Lumo ToS.
  • Political manipulation. In addition to mis- and disinformation generally, several providers included prohibitions on creating or disseminating deceptive content and false engagement or engaging in political manipulation in relation to democratic processes specifically. The ToS for ChatGPT, Claude, Ernie Bot, Copilot, Le Chat and Gemini explicitly banned interfering with elections or democratic processes.50ChatGPT UP under “empower people”; Microsoft 365 Copilot CoC-E, under “Deception, disinformation, and inauthentic activity”; Ernie Bot ToS, S.3.(2)(vii); Gemini Gen AI PUP; Claude UP, under “Do Not Undermine Democratic Processes or Engage in Targeted Campaign Activities”; Le Chat UP, under “Misinformation”.
3.6 Biometric Processing
  • Overview. Several of the ToS contained prohibitions concerning biometric processing,51ChatGPT UP under “respect privacy”; Claude UP, under “Do Not Use for Criminal Justice, Censorship, Surveillance, or Prohibited Law Enforcement Purposes”; Ernie ToS, S.3(2)(ii). that is: processing relating to the physical, physiological, or behavioural characteristics of a natural person.52See EU AI Act, Art.3(34). The relevant terms frequently reflected, and in several instances replicated almost verbatim, the prohibitions of Article 5 of the EU AI Act.53EU AI Act, Art. 5. For example, the AI Act restricts social scoring based on personal traits.54EU AI Act, Art. 5(1)(c). The ToS for ChatGPT included provisions prohibiting “social scoring” or “classification of individuals” based on protected characteristics, personal traits, or behaviour.55ChatGPT UP under “respect privacy”. The AI Act further restricts emotion inference in workplace and education institutions.56EU AI Act, Art. 5(1)(f). The ToS for ChatGPT imposed an almost identical prohibition,57ChatGPT UP under “Respect privacy”. while the ToS for Copilot and Claude prohibited the use of their services to infer emotional states in any circumstance, with Claude allowing exceptions only for medical or safety reasons.58Claude UP, under “Do Not Use for Criminal Justice, Censorship, Surveillance, or Prohibited Law Enforcement Purposes”; Microsoft 365 Copilot CoC-E, under “usage restrictions”, S.12. In addition, the ToS for ChatGPT and Copilot explicitly prohibited the use of the services to create facial recognition databases,59ChatGPT UP under “Respect Privacy”. including through the untargeted scraping of facial images from the internet or CCTV footage,60Microsoft 365 Copilot CoC-E, under “Usage Restrictions”, S.14. which maps directly to the EU AI Act.61AI Act, Art. 5(1)(e).
  • EU AI Act. The overlap suggests that some providers have attempted to align their contractual restrictions on biometric processing with EU regulations, both to anticipate compliance obligations and to show a harmonised risk posture across Member State jurisdictions. That said, some differences remained. While the EU AI Act provided detailed exceptions for law enforcement authorities,62AI Act, Article 5(1)(h): law enforcement agencies may be allowed to use real time remote biometric identification under certain conditions and for specific purposes — specifically for targeted search for specific victims, identification or localisation of a criminal suspect, or prevention of a specific, substantial, and imminent threat. most providers adopted a blanket prohibition instead. For example, Anthropic Claude’s ToS provided that the service should not be used for “Criminal Justice, Censorship, Surveillance, or Prohibited Law Enforcement Purposes”.63Claude UP, under “Universal Usage Standards”. The ToS further provided examples of prohibited practices such as “categorizing people based on their biometric data to infer their race, political opinions” and other sensitive characteristics; for criminal justice applications (including parole or sentencing); to target a person’s physical location (including through facial recognition); or for any “law enforcement application that violates or impairs the liberty, civil liberties, or human rights of natural persons”.64Ibid. The ToS for Copilot diverged in this respect, carving out exceptions that specifically permitted: (i) assessment of criminal risk when used to support human assessment of the involvement of a person in criminal activity;65Microsoft 365 Copilot CoC-E, under “Usage Restrictions”, S.9. Note that this exception is also found verbatim in EU AI Act Art. 5(1)(d). and (ii) to infer sensitive information about people by government officials, where doing so is lawful and with judicial oversight.66Microsoft 365 Copilot CoC-E, under “Usage Restrictions”, S.11.
  • No relevant provisions. By contrast, specific provisions on biometric processing were absent from the AUPs of DeepSeek, Le Chat, Lumo, MetaAI, Perplexity, and Qwen Chat. The ToS for Google’s Gemini briefly mentioned biometric data as an example of the violation of privacy rights;67Gemini Gen AI PUP, S.1(f). while the ToS for Amazon Q referred to “unlawful tracking, monitoring, and identification” as an example of the violation of privacy rights.68Amazon Q RP under “Prohibitions”.
3.7 Child Safety
  • Overview. Age restrictions were a common feature among the ToS of the GenAI services we surveyed, although the limitations and methods of application differed. Many ToS required users to be at least 13 years of age.69ChatGPT Europe ToU; Grok ToS-C under “Registration and Access”; Le Chat ToS, S.1; Lumo ToS, S.1; Meta AI UK ToS, S.1(a); Perplexity ToS, S.1(2). By contrast, Google provided a country-by-country list of minimum ages for managing a Google Account, ranging from 14 to 16 depending on the EU Member State, and set 13 as the minimum age for any unlisted countries, such as the UK.70Gemini ToS, linking to https://support.google.com/accounts/answer/1350409?hl=en-GB#zippy=%2Ceurope. Google also offered a Family Link for parents to create and manage accounts for children under the minimum age.71See https://families.google/familylink/. Other providers adopted a stricter standard, requiring users to be either 18 or the age of majority in their jurisdiction.72Amazon Q PN, under “Children’s Personal Information”; Claude Consumer ToS, S.2; DeepSeek ToS, S.2(1); Ernie Bot ToS, S.2(1); Qwen Chat ToS, S.1. The ToS for Copilot73Copilot (basic) MSA, S.4(a)(iii). based its requirements on the age of majority, but permitted parents to accept conditions on behalf of kids, thereby putting the responsibility on parents.74See also Le Chat (consumer) ToS, S.1; Grok ToS-C under “Registration and Access”. The ToS for Amazon Q similarly stated that services were not offered to children and that use under 18 required parental involvement.75Amazon Q PN, under “Children’s Personal Information”.
  • CSAM. While most ToS prohibited practices that could endanger children’s safety, other ToS (such as those for ChatGPT, Claude, xAI, and Le Chat)76ChatGPT UP under “Keep minors safe”; Le Chat UP; Claude AUP, under “Do Not Compromise Children’s Safety”; Grok AUP. went further, by committing to report child sexual abuse materials (‘CSAM’) to the National Center for Missing and Exploited Children (‘NCMEC’) or to the relevant authorities. Such terms go beyond just prohibiting harmful or illegal behaviour, by committing that providers will take affirmative action in relation to CSAM.77See also U.S. federal law, 18 U.S.C. § 2258A which requires online service providers to report suspected CSAM to the National Center for Missing and Exploited Children’s (NCMEC) Cyber Tipline once they become aware of it.
3.8 Geographic Restrictions
  • Overview. Many of the ToS we surveyed contained restrictions on the geographic availability of GenAI services. Generally speaking, the ToS adopted one of four approaches.
  • US export controls. First, some ToS prohibited the use of their services in certain countries and by certain customers based on export control provisions, with specific references to US-embargoed or sanctioned countries, entities or individuals.78See Amazon Q CA, S.11(6); ChatGPT Europe ToU under “General Terms”; Copilot Consumer ToS, S.12, under “export controls”; Grok ToS-C under “export controls”; Perplexity ToS-C, S.11(6). See further OpenAI FAQ under “supported countries”. For example, the Amazon Q ToS referred specifically to any export control laws that “apply to a US company”, including any sanctions implemented by the Office of Foreign Assets Control.79Amazon Q CA, S.11(6). The ToS for Grok contained similar wording,80Grok ToS-C under “export controls”. while the ToS for ChatGPT stated that the service must not be “exported or re-exported” to “any U.S. embargoed country or territory”.81ChatGPT Europe ToU under “General Terms”.
  • Any applicable export controls. Second, some providers referred generically to any applicable export control laws.82Copilot (basic) MSA, S.16; DeepSeek ToS, S.8; Gemini ToS, under “respect others”; Le Chat ToS, S.2(a). The ToS for DeepSeek and Le Chat restricted access to, and use of, their services based on export controls and sanctions laws.83DeepSeek ToS, S.8; Le Chat (consumer) ToS, S.2(a). The ToS for Copilot (basic) similarly required the customer to comply with “all domestic and international export laws and regulations that apply to the software and/or Services”, which included “restrictions on destinations, end users and end use”.84Copilot (basic) MSA, S.16. All three ToS referred to any applicable export control and sanctions laws, without specifying to which jurisdictions the clause referred. Moreover, while DeepSeek did not expressly restrict use in specific jurisdictions, some countries have independently chosen to ban its use within their territories. For example, Australia85The Secretary of the Department of Home Affairs Direction 001-2025 DeepSeek Products, Applications and Web Services. and South Korea86The South Korean Personal Information Protection Commission (PIPC) recommended that DeepSeek temporarily suspend its chatbot service until necessary updates are implemented to comply with applicable laws. prohibited the use of the Chinese GenAI service domestically, likely for policy reasons related to information sovereignty, privacy, or national security.
  • Supported Regions. Our survey also revealed that in addition to the two above positions, some of the above providers chose to supplement those provisions with lists of ‘supported’ or ‘unsupported regions’.87Claude Consumer ToS, under “Supported Regions Policy”. For example, the Claude ToS stated that customers could access the service only “in compliance with [...] the policy governing the countries and regions Anthropic currently supports”.88Claude Consumer ToS, under “Supported Regions Policy”. In that case we interpret the ‘unsupported region’ list to serve as a contractual ban on using the service in those regions, rather than as a disclaimer of responsibility for providing support to customers in those regions. By contrast, Google published a list of supported countries in its FAQ, but did not clearly and explicitly incorporate the FAQ into the Gemini ToS. We do not interpret the FAQ webpage as a contractual prohibition.89See the list here https://support.google.com/gemini/answer/13575153?hl=en-GB. The countries listed differed between the services that operated a list of supported, or unsupported, countries. For example, ChatGPT, Gemini, and Claude did not include China as a supported country.90Claude Consumer ToS, under “Supported Regions Policy”; Gemini API Additional Terms of Service under “Available Regions”; Open AI FAQ under “supported countries”. The ToS of Copilot did not contain an explicit prohibition on use in China. That said, a Microsoft webpage stated that Microsoft 365 Copilot was “available globally … with exceptions for regions like China (excluding Hong Kong) and embargoed markets”.91Supported regions and languages in Microsoft Copilot (last visited 1 October 2025). At the time of the survey, neither the US nor the EU had imposed general sanctions prohibiting the provision of GenAI services to users in China. Nonetheless, providers may have viewed China as a high-risk jurisdiction, because of its potential exposure to future sanctions and worries about possible abuse in conflict situations.92The U.S. Department of Justice’s Data Security Program (2025), which enforces Executive Order 14117 and limits the transfer of bulk sensitive personal data to specific high-risk jurisdictions, has recently designated China as a “country of concern”. Further, the list of supported countries for Claude did not cover Afghanistan, Mali, Yemen, Libya, or Russia — and in the case of Ukraine, made an exception for the Crimea, Donetsk, Kherson, Luhansk, and Zaporizhzhia regions.93Claude Consumer ToS, under “Supported Regions Policy”. Russia was likewise excluded from the supported-countries lists of both Gemini and ChatGPT, while Afghanistan was excluded from Gemini’s.94See Open AI FAQ under “supported countries”; Gemini FAQ, under “Where you can use the Gemini web app”.
  • Qwen Chat. Third, unlike ToS that imposed an affirmative obligation to comply with specific export control and sanctions regimes, the Qwen Chat ToS instead required users to warrant that they themselves were not subject to any “economic, financial, trade” sanctions or other restrictive measures imposed in “any jurisdiction”.95Qwen Chat ToS, S.1.2(c)(ii).
  • Ernie Bot, Meta AI, and Lumo. Finally, the Ernie Bot, Meta AI and Lumo ToS had no express export-control or geographic restrictions in their GenAI ToS.96We found that the Ernie Bot ToS provides that “The Services are currently not available for users in certain countries and regions.” However, we did not find any available lists online. One possible explanation is that both Meta and Baidu make the underlying LLMs (Llama and Ernie, respectively) available under open-source licenses (albeit with some restrictions on use), while Lumo uses underlying LLMs that are available open-source.97For a discussion of whether this means the models should strictly be considered ‘open source’ or merely ‘open weights’, see the European Open Source AI index, at https://osai-index.eu/ and M. Nolan, “Llama and ChatGPT Are Not Open-Source” (2023), https://spectrum.ieee.org/open-source-llm-not-open. This means users could deploy the underlying models directly on their own IT devices (subject to the open-source license conditions). Nevertheless, an obligation requiring the customer to comply with export controls and trade sanctions was present in ToS for Meta products other than the MetaAI service, which were outside the scope of our survey.98See Supplemental Meta Platforms Technologies Terms of Service, S.5.9.
3.9 AI Model Training
  • Overview. In this section, we look at terms that govern whether, and if so, to what extent, customers can use the outputs of a GenAI service to train or create another AI system. We observed three trends.
  • Broad prohibition. First, the ToS for Copilot (basic) strictly prohibited the use of outputs to create or train any sort of models or services at all.99Copilot (basic) MSA, under “AI Services” S.(iv). The Gemini ToS adopted a comparable position, prohibiting the use of “AI-generated content from our services to develop machine learning models or related AI technology”.100Gemini ToS, under “Don’t abuse our services”. In this context, AI-generated content appears to refer to outputs, since the services under review are generative in nature, while the prohibition on developing AI systems would likely cover “training” such systems.
  • Competing services. Second, other ToS prohibited using outputs for training or development only to create competing AI models or services. This included the ToS for ChatGPT, Claude, Ernie Bot, Grok, Le Chat, Perplexity, and Qwen Chat.101ChatGPT Europe ToU under “What you cannot do”; Claude Commercial ToS, S.D(4); Ernie ToS, S.3(2)(i); Grok ToS-E, S.2; Le Chat (consumer) ToS, S.3; Perplexity AUP, S.3(e); Qwen Chat ToS, S.II(1)(e). For example, the ToS for Grok prohibited the use of outputs to “develop models or services that compete with xAI”.102Grok ToS-E, S.2. Such terms raised a question as to what constituted a competing model. For example: would building a specialised vertical application for a specific sector count as a “competing” service, or are only general-purpose models within the scope of these prohibitions? The ToS for OpenAI’s ChatGPT Enterprise service provided a limited exception to this rule, by permitting customers to use outputs internally for classifiers or embeddings or for fine-tuning their own models, provided such models were not distributed to third parties.103ChatGPT Enterprise SA, S.17, under “permitted exception”.
  • Training permitted. Third, DeepSeek, alone among the thirteen providers we surveyed, expressly allowed the use of outputs for training other models, including ‘distillation’,104Knowledge distillation (or model distillation) is a model compression technique in which a smaller “student” model is trained to mimic the behavioural outputs of a larger “teacher” model, typically by learning from the teacher’s probability distributions or reasoning traces rather than from original labelled data. See UK Government Digital Service, AI Insights: Model Distillation (18 December 2025). as well as for “derivative product development”.105DeepSeek ToU, S.4(2). By contrast, the ToS for MetaAI, Amazon Q, and Lumo remained silent on this point.
3.10 Image and Video Generation
  • Overview. Many of the services we surveyed had the capacity to generate images or videos as output, in addition to text. The relevant AUPs we surveyed typically defined “content” broadly so as to include audiovisual materials, meaning that the general prohibitions surveyed above also applied to images and videos. Nevertheless, given the distinct risks associated with synthetic audio and visual media, some ToS contained special AUP provisions governing their creation.
  • NCII. For example, it was common for ToS to ban non-consensual intimate imagery (‘NCII’ or so-called deepfake pornography). While the exact wording varied (e.g. referring variously to “NCII”,106Microsoft 365 Copilot CoC-E under “Non-consensual intimate content”; Gemini GenAI PUP, S.1(c); Le Chat UP under “Prohibited content and activities”; ChatGPT UP. or “sexually intimate images of someone without their permission”107See Copilot content and conduct policies, under “Non-Consensual Intimate Imagery and Intimate Extortion”. See also Copilot (basic) MSA, S.4(b)(i) which incorporates the content and conduct policies.), the normative intent was clear: using GenAI systems to create such content was strictly prohibited without consent. The ToS for Le Chat, for example, explicitly stated that any intimate imagery required the express consent of all depicted individuals.108Le Chat UP. In other cases, the ToS simply prohibited the use of the service to generate any “pornographic” or “adult” content whatsoever (that is: regardless of whether any depicted individuals had consented).109See e.g. MetaAI ToS, S.1(b); DeepSeek ToS, S.3(4)(5); Ernie Bot, AUP, S.3(2)(v); Claude UP under “Do Not Generate Sexually Explicit Content”; Perplexity AUP, S.2(a); Qwen Chat ToS, S.II(2)(4).
  • Grok. By contrast, the Grok AUP prohibited “[d]epicting likenesses of persons in a pornographic manner” only insofar as doing so constituted an illegal activity.110Grok AUP, S.1(3). This implied that users could use the service to generate non-consensual intimate imagery, if doing so was not explicitly prohibited under the applicable law. In January 2026, the European Commission and the UK ICO launched investigations into Grok as a tool for generating sexualised images of real people without their consent.111L. Cress, “EU investigates Elon Musk’s X over Grok AI sexual deepfakes”, 26 January 2026, https://www.bbc.co.uk/news/articles/clye99wg0y8o/; UK ICO, “ICO announces investigation into Grok”, 3 February 2026, https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/02/ico-announces-investigation-into-grok/. Following a backlash, xAI announced that it had implemented a geo-blocking measure preventing Grok users from generating intimate images of real people “in jurisdictions where such content is illegal”.112See “@Grok Account Image Generation Updates”, https://x.com/safety/status/2011573102485127562; C. Hayes et al., “X to stop Grok AI from undressing images of real people after backlash”, 14 January 2026, https://www.bbc.co.uk/news/articles/ce8gz8g2qnlo. xAI kept its AUP the same; in effect, it supplemented its contractual prohibition on users generating such content with a technical measure preventing users from doing so in specific jurisdictions.
  • ChatGPT and Copilot. The ToS for ChatGPT and Copilot each included distinct policies specifically governing image generation.113ChatGPT Image & Video UP; Copilot Image & Video ToU. These image-generation policies broadly aligned with their general AUPs, but introduced more detailed and context-specific rules. ChatGPT’s Image and Video Policy expressly banned the use of Sora114OpenAI has an individual, model-specific product for image and video generation called Sora, in addition to the image-generation capabilities already available within ChatGPT. to generate content that recreates the likeness of living public figures without their consent or deceased public figures in contexts where the use of their likeness is not permitted.115ChatGPT Image & Video UP. We note that, prior to November 2025, ChatGPT’s policy allowed public figures to submit an opt-out request to prevent image generation of their likeness. The opt-out form was highly prescriptive and included detailed procedural steps for individuals seeking to exercise this right. However, following the policy update on 19 November 2025, this opt-out form was removed and is no longer available.
4. Termination Rights and Consequences
4.1 Customer Termination Rights
  • Overview. The ToS we surveyed typically permitted customers to terminate the contract at will. The majority of the ToS we surveyed took this approach.1See, Amazon Q CA, S.5(2)(a); Chat-GPT (basic) Europe ToS, under “Termination and suspension”; Claude Consumer ToS, S.12; Claude Commercial ToS, S.I(2); Copilot (basic) MSA, S.9(g); Microsoft 365 Copilot MCA, under “term and termination”; DeepSeek, ToS, S.2(5); Ernie Bot ToS, S.2(4); Gemini (basic) ToS, under “Suspending or terminating your access to Google services”; Gemini Workspace ToS, S.8; Grok ToS-C under “Termination or Suspension”; Grok ToS-E, S.6(2); Le Chat (consumer) ToS, S.8; Le Chat Commercial ToS, S.11(2); Meta Products ToS, S.4(1); Perplexity (consumer) ToS, S.2(3). This made sense, since contracts for GenAI services were typically of an indefinite duration,2See e.g. Amazon Q CA, S.5(1); Claude Commercial ToS, S.I(1); Microsoft 365 Copilot MCA, under “Term and termination”; Grok ToS-E, S.6(1); Le Chat ToS, S.8 (consumer). meaning they will continue unless either party terminates. For example, the Gemini (basic) ToS stated: “you’re always free to stop using our services at any time”.3Gemini (basic) ToS, under “Suspending or terminating your access to Google services”. The Gemini Workspace ToS similarly allowed customers to “terminate this Agreement for [...] convenience at any time on prior written notice”.4Gemini Workspace ToS, S.8(5). The ToS for Grok likewise stated that customers could “stop using our Service at any time and close your account”.5Grok ToS-C, under “Termination, Suspension, Discontinuation”.
  • Lumo and Perplexity. By contrast, the Lumo and Perplexity (consumer) ToS did not expressly grant customers the right to terminate their subscription at any time. Instead, the Lumo ToS explained that the service was provided on a monthly, yearly, or bi-yearly basis, with automatic renewal unless cancelled by the user. This implied that a customer’s termination rights depended on the service term for which it had signed up (e.g. customers on a monthly subscription could cancel at the end of the month, etc.).6Lumo ToS, S.4. Perplexity’s ToS stated that “To avoid future subscription charges, you must cancel your subscription before the subscription period renewal date”.7Perplexity Consumer ToS S.2(3).
  • ChatGPT Enterprise. Further, the ToS for ChatGPT Enterprise stated that the contract was for a defined term (as agreed between the parties) and only permitted the customer to terminate the contract within that term for cause (namely a material breach not cured within 30 days, a cessation of business operations, or insolvency).8Chat-GPT Enterprise, S.11(1) and (2). Customers could also terminate the contract if a product update “materially affects functionality”, provided they gave 30 days’ written notice within five business days of receiving the change notice.9Chat-GPT enterprise SA, S.2(3) 5(1). The ToS of Perplexity (commercial) adopted a similar approach, whereby commercial customers could only terminate the agreement for breach of contract if the breach was not cured within 30 days.10Perplexity (commercial) ToS, S.9.3. The DeepSeek ToS were more difficult to interpret, since they permitted customers to terminate the contract “under the agreed conditions”, without specifying what those conditions were.11DeepSeek, ToS, S.2(5).
  • Qwen Chat. Lastly, the ToS for Qwen Chat did not contain an express provision permitting customers to terminate at will or in case of material breach by the provider. Instead, the ToS stated only that the customer could terminate if Alibaba changed the terms of the contract, in which case the customer “must immediately instruct us in writing to terminate your account and you must immediately cease to access and/or use our Services”.12Qwen Chat ToS, under “Thank you for using Qwen Chat!”. Presumably, customer termination in all other circumstances would be governed by the general rules applicable under the law chosen in the contract (Singapore), and any applicable rules of European consumer protection law which might grant a contracting party a right to terminate in case of a material breach by the other party.
  • Consumer v. commercial use. Some termination clauses applied different notice periods to consumer and commercial use. For example, the ToS that governed consumer use of Claude allowed consumers to terminate without observing a notice period,13Claude Consumer ToS, S.12, under “termination”. while the ToS that governed commercial use set a 30-day notice period for customer termination.14Claude Commercial ToS, S.I(2). The Copilot ToS also differentiated between consumers, who could terminate immediately,15Copilot (basic) ToU under “important disclosures & warnings”; Copilot (basic) MSA, S.9(g). and commercial customers, who had to provide 60 days’ notice when terminating without cause.16Microsoft 365 Copilot MCA, under “term and termination”. The ToS for Grok, by contrast, permitted both consumers and commercial customers to terminate immediately.17See Grok ToS-C, under “Termination or Suspension”; Grok ToS-E, S.6.
  • For cause and at will. In some cases, the ToS contained clauses that permitted the customer to terminate the contract for cause as well as at will. For example, the ToS for Amazon Q, Claude (commercial), Microsoft 365 Copilot, Gemini Workspace, Grok (commercial) featured a separate termination for cause provision.18Amazon Q CA, S.5(2)(b); Chat-GPT Enterprise SA, S.11; Claude Commercial ToS, S.I(2); Microsoft 365 Copilot MCA, under “term and termination”; Gemini Workspace ToS, S.8(3)(b); Grok ToS-E, S.6(2). The terms gave the customer the right to terminate the contract in the event of a material breach that remained uncured for 30 days following written notice.19Ibid. This notice period was the same as for termination at will (summarised above), but permitted the provider to cure the breach within the notice period to avoid termination. By contrast, the Microsoft 365 Copilot ToS notice period for cause was shorter than the 60-day period for termination at will: the ToS permitted commercial customers to terminate for cause if a material breach remained uncured for 30 days following written notice.20Microsoft 365 Copilot MCA, under “term and termination”.
4.2 Provider Termination Rights
  • Overview. Across the ToS we surveyed, we found three different approaches to provider termination rights.
  • Termination at will — commercial use. First, four of the thirteen providers had ToS which granted the provider a right to terminate at will in relation to commercial customers. The ToS for Amazon Q and Claude (commercial), and Microsoft 365 Copilot all permitted the provider to terminate at will, by providing either 30 days’21Amazon Q CA, S.5(1); Claude Commercial ToS, S. I(2). or 60 days’ notice.22Microsoft 365 Copilot MCA, under “term and termination”. Further, the terms that governed commercial use of the Grok service permitted immediate provider termination “at any time, and at its sole discretion”.23Grok ToS-E, S.6. These terms differed from those that governed consumer use, under which providers could only terminate the contract for cause (for instance, in the case of customer material breach of the terms, non-payment, a legal requirement, the risk of harm to the provider, or inactivity in case of free accounts).24Claude Consumer ToS, S.12 under “Termination”; Grok, ToS-C, under “Termination or Suspension”; Copilot (basic) MCA, S.4(a)(ii), 4(b)(iv), 9(a). The Amazon Q service is aimed solely at commercial use, so there are no equivalent terms for consumer use.
  • Termination for cause. Second, six of the thirteen providers only had the right to terminate for cause, in relation to both commercial and consumer use. The ToS for ChatGPT, DeepSeek, Gemini, Le Chat, Lumo and MetaAI only permitted the provider to terminate for listed grounds, which included customer material breach of the terms, non-payment, a legal requirement, or the risk of harm to the provider.25Chat-GPT ToS under “Termination and suspension”; ChatGPT Enterprise SA, S.11; DeepSeek, S.7(2); Gemini (basic) ToS, under “Suspending or terminating your access to Google services”; Gemini Workspace ToS, S.8(3); Le Chat ToS, S.8; Le Chat Commercial ToS, S.11(3); MetaAI ToS, S.3. These terms applied equally to commercial and to consumer use.
  • Termination at will — consumer use. Third, by contrast, the ToS for Qwen Chat (which was for non-commercial use only)26Qwen Chat ToS, S.II(3). and Perplexity (consumer) granted the providers the right to terminate a consumer’s account respectively at “Alibaba’s sole discretion”27Qwen Chat ToS S.I(2) and (4). and “for any or no reason”.28Perplexity Consumer ToS, S.11(3). Both ToS did not mention any notice period in relation to such termination, which suggested that the termination could be immediate. Further, the ToS for the Ernie Bot service (which was for consumer use only) stated only that the provider would “make reasonable efforts to notify you in advance”, while reserving the right to terminate the agreement “at any time” and “for any reason”, “without prior notice”.29Ernie Bot ToS, S.5(1), 10(1).
  • Notice periods. Where a provider could terminate for cause, it was typical for providers to commit to notify the customer that they were in material breach of the agreement first and offer a 30-day period for the customer to cure the breach.30Amazon Q CA, S.5.2; Chat-GPT Enterprise SA, S.11; Claude Consumer ToS, S.12, under “termination”; Claude Commercial ToS, S.I(2); Microsoft 365 Copilot MCA, under “term and termination”; Gemini Workspace ToS, S.3(4); Grok ToS-E, S.6; Perplexity (commercial) ToS, S.9(3). This was similar to the 30-day notice periods offered by providers who had a right to terminate commercial customers at will. That said, several ToS also permitted immediate termination without a cure period or a notice period in case of (1) compliance or security risks;31Claude Consumer ToS, S.12, under “termination”; Ernie ToS, S.10(1); Grok ToS-E, S.5(3); Gemini (basic) ToS, under “Suspending or terminating your access to Google services”; Le Chat (consumer) ToS, S.8. (2) where necessary to comply with legal obligations or to respond to unlawful activity;32Amazon Q CA, S.5(2); Chat-GPT Europe ToS, under “Termination and suspension”; Microsoft 365 Copilot MCA, under “term and termination”; Copilot (basic) MSA, S.4(b)(iv); Gemini (basic) ToS, under “Suspending or terminating your access to Google services”; Gemini Workspace ToS, S.8(6); Grok ToS-E, S.5(3); Lumo ToS, S.(4); Perplexity (commercial) ToS, S.9(3). or (3) in response to an immediate risk of harm to the company or third parties.33Chat-GPT Europe ToS, under “Termination and suspension”; Gemini (basic) ToS, under “Suspending or terminating your access to Google services”; Qwen Chat ToS, S.I(4). Additional triggers appeared in other ToS, including if a third-party technology dependency expired;34Amazon Q CA, S.5.2. or if the customer ceased operations or entered insolvency.35Chat-GPT Enterprise SA, S.11(2); Grok ToS-E, S.6(2).
  • Analysis. For an analysis of provider termination rights under UK and European consumer protection law, download our full paper free from SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698.
4.3 Post-Termination Data Retrieval
  • Overview. We found significant differences in how (and whether) the ToS granted customers a right to retrieve data following termination.
  • Data retrieval period. Four of the 13 providers specified a formal grace period for data retrieval. Amazon allowed customers a 30-day period after termination to retrieve their content, provided all outstanding fees had been paid.36Amazon Q CA, S.5(3)(b). The Claude (commercial) ToS had a similar 30-day window, during which customers may either request a copy of all data or use a self-service retrieval feature.37Claude DPA, S.H(1)(a). That said, the Amazon Q provision did not apply in case the provider terminated the agreement for cause, including because of an uncured material breach by the customer or in case of a provider legal obligation.38Amazon Q CA, S.5(3)(b). The Microsoft 365 Copilot ToS stated that Microsoft would retain customer data for 90 days after termination, so the customer can extract data.39Microsoft 365 Copilot DPA, under “Data Retention and Deletion”. The ToS for ChatGPT Enterprise merely stated that, upon termination, the provider would “at Customer’s instruction, return or delete Customer Data”, albeit without specifying a time limit for the retrieval.40ChatGPT Enterprise DPA, S.2(11).
  • Consumer v. commercial use. By contrast, the terms that governed consumer use of the same services were less favourable to customers. For example if Microsoft terminated the contract under the Copilot (basic) ToS, the customer’s “right to access the Services stops immediately” and Microsoft will “delete Data or Your Content”, unless required by law to keep or return it.41Copilot (basic) MSA, S.4(a)(iv). The ToS for Claude (consumer) similarly stated that, upon termination, the provider “may at our option delete any Materials or other data associated with your account”.42Claude consumer ToS, S.12. The ToS for ChatGPT (basic) stated that if OpenAI terminated a customer account, it would “make reasonable efforts to notify you in advance so you can export your Content or your data from the Services”, albeit without providing any particular timeframe for retrieval.43Chat-GPT Europe ToS, under “Termination and suspension”.
  • Grok and Perplexity. By contrast, the ToS for Grok (commercial) and Perplexity (commercial) did not contain explicit post-termination data retrieval rights. Instead, xAI stated that it would “return [...] or destroy” confidential information (which included customer content) within 30 days of termination.44Grok ToS-E, S.6(3) and 10(1). The ToS for Perplexity (commercial) contained a similar term.45Perplexity ToS-E, S.9(5). This left open the possibility that a provider might destroy customer data post-termination, without returning a copy to the customer.
  • No relevant provisions. The remaining nine of the thirteen providers (including Grok and Perplexity) did not include explicit post-termination data retrieval provisions in their ToS.46See the ToS for Gemini, DeepSeek, Ernie Bot, Grok (consumer), Le Chat, Lumo, Meta AI, Perplexity (consumer) and Qwen Chat.
  • Limitations. Finally, the data retrieval processes covered customer data, which would likely include past session data. However, no provider surveyed explicitly addressed the export of model derivatives such as fine-tuned weights, classifiers, or trained configurations. As a result, customers who have fine-tuned or re-trained a GenAI model might not be able to retrieve such data post-termination.
4.4 Post-Termination Data Erasure
  • Overview. The ToS of the providers we surveyed adopted different approaches to post-termination data deletion. Commercial agreements typically provided greater clarity and procedural detail than consumer ToS. The ToS for ChatGPT Enterprise, Claude (commercial), and Microsoft 365 Copilot included explicit commitments to delete customer data within a certain timeframe (such as 30 days after termination).47Chat-GPT enterprise SA, S.11(3); Claude DPA, S.H(1); Microsoft 365 Copilot DPA, under “Data Retention and Deletion”; Grok ToS-E, S.6(3). For example, the terms governing Microsoft 365 Copilot stated that Microsoft would delete customer data 180 days after termination.48Copilot (commercial) DPA, under “Data Retention and Deletion”. The Amazon Q ToS stated that “[f]ollowing closure of your AWS account, we will delete Your Content in accordance with the technical documentation applicable to the Services”.49AWS ST, S.1. The ToS itself does not specify a deletion timeframe, instead deferring to external documentation, the content and precision of which may vary by service. The Lumo ToS stated that the provider would “delete all Company Personal Data”, without committing to any particular timeframe for doing so.50Lumo DPA, S.9.
  • Grok and Le Chat. Further, the Grok (commercial) ToS stated that “each party will return to the other party (or destroy) such other party’s Confidential Information [...] within 30 days”, which included user content.51Grok ToS-E, S.6(3). Similarly, the DPA for Le Chat stated that upon termination, Mistral would “delete or return” all personal data processed on behalf of the customer “in accordance with Mistral AI’s deletion policies and procedures”.52Le Chat (commercial) DPA, S.10(1). These terms left some uncertainty, since the provider did not explicitly commit to deleting any remaining copies of data from its servers, if it had returned a copy of such data to the customer.
  • Consumer v. commercial use. By contrast, the ToS for ChatGPT (basic), Claude (consumer), Copilot (basic), and Grok (consumer) contained either no post-termination deletion provisions or only vague ones.53For brief post-termination deletion provision see, Claude Consumer ToS, S.12 under “termination”; Copilot MSA, S.4(a)(iv). For example, the terms that governed consumer use of Copilot (basic) stated only that Microsoft would delete user content associated with closed accounts, unless the law required its retention or transfer, without setting a particular timeframe for deletion.54Microsoft 365 Copilot MCA, under “term and termination”. Further, the Perplexity ToS (consumer) expressly state that the company “may, but is not obligated to, delete any of Your Content” and “shall not be responsible for the failure to delete or the deletion of Your Content”.55Perplexity (consumer) ToS, S.11(3). Thus, consumers have no guarantee that the provider will delete their data after termination.
  • No relevant provisions. Lastly, the ToS for DeepSeek, Ernie Bot, Qwen Chat, Le Chat (consumer) and Meta AI did not include explicit deletion terms. In these cases, termination led to account closure, but the fate of customer data remained contractually undefined, leaving customers uncertain about how long copies of their data may persist on the providers’ servers. The Meta AI ToS only stated that customers should visit its Privacy Center webpages for instructions on how to delete personal data from the service during use.56Meta AI UK ToS, S.2(C). However, the term did not include an explicit provider commitment to delete customer data post-termination, and a customer would be unlikely to have the ability to delete its data from the service once their account was closed.
Methodology
Selection of Services
  • We surveyed 20 terms of service ('ToS') that applied to generative Artificial Intelligence ('GenAI') services. We selected services that featured a chat-based interface through which the user inputs prompts. The prompts are then processed by a Large Language Model ('LLM') that generates an output. We focused on general-purpose GenAI services which generate outputs in the form of text and images, as opposed to specialised services which focus on generating a particular type of content (such as images or video)1Examples include Midjourney, Adobe Firefly, Google Veo3, and Kling. or on a particular sector.2Examples include Harvey AI for the legal sector; or FinLLM for financial services. Thus, the scope of our survey closely tracked the definition of a 'general-purpose AI system' under the EU AI Act, that is: a system based on an AI model that "displays significant generality and is capable of competently performing a wide range of distinct tasks".3EU AI Act, Art.3(63) and (66), Recital 97.
  • We limited our review to GenAI services which allow users to interact with LLMs through conversational chat-based interfaces (e.g. a web chat portal or dedicated application). These services might include advanced features that allow users to fine-tune the underlying LLM for specific tasks or sectors, such as by re-training the model with a task-specific dataset or through human feedback.4For more information on fine-tuning, see H. Chawre, "Fine-Tuning LLMs: Overview, Methods, and Best Practices", (2025), https://www.turing.com/resources/finetuning-large-language-models. They might also include tools to embed the chat-based AI functionality within the user's own systems and integrate it into the user's processes. By contrast, we did not cover services in which a provider gives customers direct, non-conversational access to an LLM (e.g. through APIs, software development kits, or by allowing the customer to host the model themselves, whether on-premises or in the cloud). So, for example, we included Meta's MetaAI service in our survey, but did not include Meta's provision of access to the underlying Llama LLM.5The same applies to OpenAI's ChatGPT services and its GPT OSS models and DeepSeek's chat-based service and its open-source model.
  • We based our selection of 13 providers on two criteria: (1) market share and (2) regional spread. As to the first criterion, we aimed to survey those GenAI services which customers in the UK and the EU use most. That said, since GenAI services are a new type of service, market shares are hard to assess. Further, market shares in such an emerging market are likely to be both fragmented and dynamic. Given the absence of reliable market share data, our selection of services necessarily involved a subjective element, as informed by news reports and anecdotal evidence. Second, we aimed to include services offered by providers who are headquartered in different regions, to see if the ToS differed depending on the providers' location. To that end, we selected eight providers headquartered in the US; three headquartered in China; and two headquartered in Europe, as set out in the table at the top of this page. Providers generally offered their GenAI services based on a 'freemium' model, comprising a basic service layer available for free, with optional paid service plans. These paid service plans were offered at different prices and with different features, often marketed as 'Pro', 'Team', 'Plus', or 'Enterprise' plans. In practice, both the free and paid versions of a service were typically covered by the same ToS, although specific clauses within the contract might differ between free and paid use of the service.
  • We looked at services used by both consumers (that is, for personal, non-commercial use) and business users (for commercial use, defined broadly).6We use the terms 'business' and 'commercial' in a broad sense here to mean non-consumer. This also covers non-commercial organisations (such as public sector bodies and charities), as well as individuals acting in a professional capacity. We use the term 'consumer' to refer to a natural person acting for purposes wholly or mainly outside of their trade, business, craft, or profession. This definition follows European consumer protection law. See e.g. Council Directive 93/13/EEC of 5 April 1993 on unfair terms in consumer contracts, Art.2(b). Some services were positioned as being for use by only one or other of these groups. For example, the ToS for Alibaba Cloud's Qwen Chat service specified that the service was for personal, non-commercial, use only,7Qwen Chat ToS, S.II(3). whereas the Amazon webpage promoting the Amazon Q service made clear that the service was intended for business use only. Amazon promoted Amazon Q as a "generative AI assistant that transforms how work gets done in your organization" and "helps every employee get insights on their data and accelerate their tasks".8See https://aws.amazon.com/q/ [visited on 14 October 2025].
  • Other services were intended for both consumer and commercial use. This complicated the survey, because many providers had different ToS documents, or at least some alternate provisions, depending on whether the customer used the service for consumer purposes, or for commercial purposes. For example, the Claude, Perplexity, and Grok services had separate ToS documents for consumer and for commercial use.9Claude Consumer ToS; Claude Commercial ToS; Grok ToS-C; Grok ToS-E; Perplexity ToS-C, S.5(1); Perplexity, ToS-E. Other providers, such as Meta, had a single set of ToS documents that governed both consumer and commercial use of the service, but which contained specific clauses that applied only to consumers.
  • For ease of reference, in our findings below, we refer simply to the 'Claude ToS' or the 'Grok ToS' where the terms are in substance the same for consumer and commercial use. Where the terms differ in substance, we distinguish between the terms that govern commercial and consumer use of a service as ToS (commercial) and ToS (consumer). For example, we refer to the ToS of Claude (commercial) and Grok (consumer). We use this approach, regardless of whether the clauses are contained in separate ToS documents (as with Claude and Grok), or merged into the same set of ToS documents (as with MetaAI and Lumo). This allows us to focus on the substance of the terms, irrespective of the provider's approach to dividing up its ToS documentation.
  • The situation was more complicated for three providers, namely: Google, Microsoft, and OpenAI. Each provider offered different services, depending on whether the customer signed up as an individual user or as an enterprise customer. For example, OpenAI offered a general ChatGPT service, which was aimed primarily at individual customers. Customers could use this for personal or commercial purposes. We refer to this service as ChatGPT (basic). The ChatGPT (basic) service was governed by a set of ToS documents that contained different clauses that applied to consumer and commercial use. Separately, OpenAI offered the ChatGPT Enterprise and ChatGPT Business services, which were aimed primarily at larger organisations.10See "Introducing ChatGPT Enterprise", https://openai.com/index/introducing-chatgpt-enterprise/; "ChatGPT for Business", https://chatgpt.com/business/business-plan/. These services had their own ToS.11ChatGPT Europe ToU; ChatGPT Enterprise OpenAI SA. For simplicity, we focused on 'ChatGPT Enterprise'. This meant OpenAI had two sets of ToS documents that contained clauses for three different kinds of user: the ChatGPT (basic) ToS covered both consumer use and commercial use of ChatGPT (basic); while the ChatGPT Enterprise ToS covered the use of ChatGPT Enterprise by larger organisations.
  • Microsoft and Google also offered their Copilot and Gemini services on different terms, depending on how the customer accessed the service. On the one hand, customers could sign up to a basic version of the Copilot and Gemini services directly, such as through a web portal or by downloading an application. On the other hand, customers could take Copilot and Gemini services as part of a bundle with other Microsoft 365, Azure, Google Workspace, or Google Cloud Platform services. This made it difficult to determine the exact set of terms documents that applied to each use case.
  • In Microsoft's case, this was probably due to Copilot's integration with Microsoft's wide range of existing Software-as-a-Service productivity services (including Microsoft 365), which have their own complex licensing arrangements.12See e.g. the services listed at https://www.microsoft.com/licensing/terms/product/PrivacyandSecurityTerms/ and the different terms available at https://www.microsoft.com/licensing/terms/product/ForOnlineServices/. For simplicity, we focused on two sets of terms: the Copilot (basic) ToS and the Microsoft 365 Copilot ToS.13Microsoft offered "Microsoft Copilot", "Microsoft 365 Copilot", and "Microsoft 365 Copilot Chat" as separate services. We understand individual use of "Microsoft Copilot", when based on an individual's personal Microsoft Account linked to their personal email address, to be governed by the Copilot AI Experiences Terms and the Microsoft Services Agreement (and the other documents incorporated by reference therein). By contrast, most commercial use of "Microsoft 365 Copilot Chat" and "Microsoft 365 Copilot" will be based on an individual's so-called 'Microsoft Entra ID' which is linked to their corporate email address. In that case, use of the service is governed by the Microsoft Product Terms (including the Microsoft Customer Agreement and the Privacy and Security Terms) and the Data Processing Addendum. The Copilot (basic) ToS stated that the service was for a customer's "own personal use" only.14Copilot MSA, S.13(r); Copilot ToS. Further, if a user used Copilot as part of a Microsoft 365 Personal or Family service, this was also only for "personal, non-commercial use", see Copilot MSA, S.13(h). We understood this to mean the customer could only use the service for consumer (i.e. non-commercial) purposes. By contrast, the Microsoft 365 Copilot ToS was intended for business users. We therefore understand the Copilot ToS to differentiate between two types of users: consumer users under the Copilot (basic) ToS and commercial users under the Microsoft 365 Copilot ToS. By contrast, we did not cover the ToS that governed Microsoft Azure Copilot.15See https://azure.microsoft.com/en-us/products/copilot.
  • In Google's case, Google offered different versions of Gemini that were integrated either with its Google Workspace or Google Cloud Platform offerings.16See e.g. the services listed at https://docs.cloud.google.com/gemini/docs/overview, https://support.google.com/a/answer/13623623 and https://docs.cloud.google.com/gemini/enterprise/docs/editions. For simplicity, we focused on two sets of terms: the Gemini (basic) ToS and the ToS that governed the use of Workspace with Gemini (which we refer to as 'Gemini Workspace').17We understand the use of Gemini (basic) to be covered primarily by the general Google Terms of Service at https://policies.google.com/terms?hl=en; while the use of Gemini Workspace is covered primarily by the Google Cloud Terms of Service at https://workspace.google.com/terms/premier_terms/. The Gemini (basic) ToS contained clauses both for consumer and for commercial use. So, as with OpenAI above, this meant we covered two sets of ToS documents that governed three types of users: one set of general terms that applied to both consumer use and to commercial use of Gemini (basic); and one for enterprise use of Gemini Workspace. We did not cover the ToS for use of Gemini as part of the Google Cloud Platform, nor its separate Gemini Enterprise service (which offered an agentic AI platform).18See https://cloud.google.com/gemini-enterprise?hl=en.
  • Lastly, we focused on the terms that applied to customers in the UK and in the EU. Providers typically offered a separate set of ToS depending on the customer's location. For example, providers often had different terms which applied to customers in the US and to customers in Europe. In a small number of cases, providers also had specific ToS documents for individual countries within Europe, such as tailored terms for French or German customers. For example, Mistral had separate ToS for users of its Le Chat service in France and Italy;19See ToS available here: https://mistral.ai/fr/terms#terms-of-service and https://mistral.ai/it/terms#termini-di-servizio. while Google had separate ToS documents for users of its Gemini service in various European countries.20See for example, the "country versions" of Google's ToS for Germany, France, and Ireland at https://policies.google.com/terms?hl=en-GB&gl=DE; https://policies.google.com/terms?hl=en-GB&gl=FR; https://policies.google.com/terms?hl=en-GB&gl=IE. In other cases, providers offered a unified set of ToS which applied irrespective of the customer's location, but which contained specific clauses that applied to customers in particular locations. In our findings below, we summarise the terms that applied to customers in the UK and in the EU and highlight differences between the two where relevant.21For the analysis of Google's European ToS, we compared the terms that applied to users in Germany, France, and Ireland as indicative of the terms that applied across Europe. In general, we cite the terms that applied to users in Germany in the footnotes. Where relevant, we highlight any differences between the German, French, and Irish ToS. We did not analyse the terms that applied to customers in the US or elsewhere, nor the terms that applied to customers in specific EU Member States. Using these criteria, we identified 20 different sets of terms of service documents for the GenAI services offered by the 13 providers we selected, as set out in the Table of Terms.
Data Collection
  • To gather data, we accessed the ToS as stored on the providers' websites in April 2025. We used this dataset to conduct an initial analysis of the terms. However, providers amended their ToS frequently in the months that followed. For example, the Le Chat terms were updated on 27 May 2025; the ChatGPT Enterprise terms on 31 May 2025; and the Grok Enterprise terms on 9 June and again on 27 June 2025. To keep our findings up-to-date, we accessed the provider's websites to capture the updated versions of the ToS in January 2026. The findings below are based on the ToS available as of mid-January 2026 (but, unless stated explicitly, will not reflect any changes made after that). For the latest versions, please consult the providers' websites directly.
  • Given this methodology, the survey was restricted to standard form, publicly available ToS. Some customers, such as large companies or governments, may negotiate separate contractual terms with providers, tailored to fit their specific requirements. Any such negotiated terms were not covered by this survey. Further, we focused exclusively on the relationship between the GenAI provider and the GenAI customer. We did not review any contracts that applied elsewhere in the supply chain (such as between a GenAI provider and any underlying cloud infrastructure provider).
  • The ToS we surveyed ranged from the relatively short and simple to the lengthy and complex. Many ToS comprised multiple documents that were incorporated by reference, which together made up the full agreement. The types of documents often included:
    1. A general Terms of Service or Terms of Use document. This set out the overall relationship between the customer and the provider and set out commercial terms (including the parties' duties, rights, and remedies in case of breach) and covered general legal issues such as choice of law and forum. It typically incorporated other documents by reference.
    2. An Acceptable Use Policy ('AUP') which limited the ways in which customers could use the service.
    3. A Privacy Policy which governed the provider's processing of personal data and compliance with applicable data protection laws such as the General Data Protection Regulation ('GDPR').
    4. A Data Processing Addendum ('DPA') which governed the provider's role and responsibilities as a processor, when processing personal data on behalf of the customer who acted as a controller under the GDPR.
  • We included in our review all such documents incorporated by reference into the ToS. That said, since we did not set out to cover data protection clauses, most of the information in the provider's privacy policies and DPAs was not relevant to our current survey. We reference specific sub-documents in the footnotes using the short forms set out in the Table of Terms.
  • Some providers had a separate set of terms applicable to the use of their website (as opposed to the use of their GenAI services). Many providers also offered further information on the use of their services through frequently-asked-questions pages ('FAQs') or help centres, none of which were explicitly incorporated by reference into their ToS.22In some cases, it was not entirely clear whether the provider intended to incorporate a webpage into its ToS, as when a clause in the ToS referred the customer to a hyperlink "for more information". For example, the Privacy Policy for OpenAI's ChatGPT service stated that: "For information about how we collect and use training information to develop our language models that power ChatGPT and other Services, and your choices with respect to that information, please see this help center article", while linking through to a webpage that formed part of its FAQ, namely: "How ChatGPT and our foundation models are developed", https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed. As a result, it was unclear whether OpenAI intended this page of its FAQ to form part of the contract. Finally, providers typically used cookies and other tracking and customisation technologies in their websites and services, often covered by cookie policies. We did not cover separate website-specific ToS, FAQs, or cookie policies in our survey. Finally, providers often have separate (and more restrictive) clauses that govern optional third-party applications or services which customers can use in conjunction with the provider's GenAI service,23See e.g. Copilot MCA, under "Non-Microsoft Products"; Copilot MSA, S.5; Amazon Q AWS ST, S.50(10). as well as so-called 'beta' or 'pilot' features or versions of the service.24See e.g. Amazon Q AWS ST, S.2; ChatGPT OpenAI SA, S.12(3). These separate terms are out of the scope of our current survey, which focuses on the provider's main GenAI service.
Table of Terms
The full paper includes a table listing the key Terms of Service (ToS) documents cited in our survey, with a link to the web version of the current ToS. Links to webpages are accurate as of 1st March 2026. The original ToS as surveyed in April 2025 and in January 2026 are on file with the authors. The table is in the Annex to our paper which is available for free download from SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6350698.
About the Authors
  • Cloud Legal Project. This paper has been produced by members of the Cloud Legal Project, Centre for Commercial Law Studies, Queen Mary University of London. The authors are grateful to Microsoft for the generous financial support that has made this project possible. Responsibility for views expressed, however, remains with the authors. Find out more about the project here: www.qmul.ac.uk/cloudlegal.
  • Johan David Michels is a Researcher and PhD Candidate, Cloud Legal Project, Centre for Commercial Law Studies, Queen Mary University of London.
  • Christopher Millard is Professor of Privacy and Information Law and Project Leader, Cloud Legal Project, Centre for Commercial Law Studies, Queen Mary University of London and Senior Counsel, Bristows LLP.
  • Camil Abou Farhat is an Assistant Researcher, Cloud Legal Project, Centre for Commercial Law Studies, Queen Mary University of London and Data Protection Advisor, Bird & Bird.
Download the full paper free from SSRN →