1. Which law and which court?
Your contract is only as good as the forum you can enforce it in. A PRC-law contract with disputes in Beijing is a different risk from Ontario or California law, whatever the model quality.
Compare
The comparison that decides a deployment: terms, not benchmarks.
Governing law, whether your prompts train the model, where data lives, how a copyright notice is handled, and which governments have restricted the product. Read from each provider’s published terms as of 18 October 2026; corrections welcome through the contact form.
Side by side
Eleven assistants on five legal dimensions
| Product | Operator and governing law | Consumer prompts used for training? | Data location | Copyright and DMCA | Government restrictions |
|---|---|---|---|---|---|
| ELP GPT | ELP Ventures (New York, US); engineering in Oakville, Ontario Province of Ontario and federal laws of Canada; informal resolution first | ELP does not use conversations to train an ELP model without separate notice and consent; some requests are processed by independent third-party providers under their terms, disclosed in the Privacy Policy | Database and storage on managed infrastructure in North America; provider processing may be elsewhere; Business/Enterprise can restrict providers and regions | Published notice-and-takedown procedure (DMCA §512 and Canadian notice-and-notice) with a designated contact, counter-notice and repeat-infringer policy; IP indemnity available in Enterprise agreements | None known |
| ChatGPT (OpenAI) | OpenAI, L.L.C. (US); OpenAI Ireland Ltd for EEA/UK/Switzerland US consumers: California law with arbitration; EEA/UK: local consumer law applies | Consumer plans: content may be used to improve models unless you turn off "Improve the model for everyone"; Team, Enterprise and API: not used by default | US-based with regional options for business tiers; EU data residency offered for enterprise/API | Copyright complaint process published; "Copyright Shield" indemnity for Enterprise and API customers (announced Nov 2023) | Italian DPA fined OpenAI €15M (Dec 2024) over GDPR; temporary Italian suspension in 2023; various public-sector usage policies |
| Claude (Anthropic) | Anthropic, PBC (US); Anthropic Ireland for EEA Consumer terms: California law; commercial terms negotiated | Free/Pro/Max: since Sept 2025 a training setting (default on in the prompt) with retention up to five years when on, 30 days when off; Team, Enterprise, API and cloud marketplaces excluded | US-based; enterprise and API customers can use regional cloud offerings | Copyright notice process published; IP indemnity in commercial terms for paying business customers | None widely reported for government device use |
| Gemini (Google) | Google LLC (US); Google Ireland Ltd for EEA US: California law; country-specific terms elsewhere | Consumer Gemini Apps: activity may be reviewed and used to improve services unless activity is turned off; Workspace and Cloud customers: not used to train without permission | Global infrastructure; Cloud customers choose regions | DMCA process published across Google; generative-AI indemnity for Google Cloud customers (training-data and output indemnities, Oct 2023) | Product-level restrictions by country; no broad government bans reported |
| Copilot (Microsoft) | Microsoft Corporation (US); Microsoft Ireland for EEA US: Washington State law; local terms elsewhere | Consumer Copilot: conversations may be used to train unless opted out in settings; Microsoft 365 Copilot and Azure OpenAI: not used for training | Global with EU Data Boundary for commercial cloud | DMCA process published; Customer Copyright Commitment indemnifies commercial customers using built-in guardrails (Sept 2023) | Some legislatures restricted staff use of consumer Copilot in 2024; commercial product widely approved |
| Grok (xAI) | xAI Corp. (US) US; see published terms for forum and arbitration | Conversations and X posts may be used to train unless opted out in X settings; business API terms differ | US-based | Copyright complaint channel through X/xAI terms; no public indemnity programme widely reported | Irish DPC opened an inquiry into training on EU users’ X data (2025); several regulators questioned image-generation safeguards |
| Meta AI (Meta) | Meta Platforms, Inc. (US); Meta Platforms Ireland for EEA US: California law; local terms elsewhere | Interactions with Meta AI and public content may be used to train; EU/UK users had an objection process; no general consumer opt-out in the US | Global infrastructure | DMCA process published across Meta products; no consumer indemnity | EU training paused in 2024 after Irish DPC engagement and resumed in 2025 with objection rights; ongoing litigation over training data |
| Le Chat (Mistral AI) | Mistral AI SAS (Paris, France) French law; GDPR applies | Consumer Le Chat: can be used to improve unless the setting is turned off; commercial and API: not used without agreement | EU-hosted; a natural fit for EU residency requirements | Notice process under French and EU law; indemnity negotiated in enterprise contracts | None reported |
| DeepSeek (Hangzhou DeepSeek) | Hangzhou DeepSeek Artificial Intelligence Co. (China) People’s Republic of China law | Inputs and outputs may be used to improve services per the privacy policy; open-weight models can be self-hosted under permissive licences, which removes this concern for self-hosters | Consumer service stores data on servers in the PRC per its privacy policy | No DMCA-style designated agent published; complaints by email under PRC procedures | Restricted or banned on government devices in Italy, the Netherlands, Taiwan, the Czech Republic, Australia, New York State and others (2025); South Korea suspended downloads then allowed them after changes; EU and US investigations ongoing |
| Qwen (Alibaba) | Alibaba Cloud: Singapore entity for international customers; Chinese entities for mainland services International: Singapore law in Alibaba Cloud international terms; mainland: PRC law | Alibaba Cloud Model Studio states customer data is not used to train without consent; consumer Qwen app terms differ; open-weight Qwen models can be self-hosted (Apache 2.0 for many) | Customer-selected regions on Alibaba Cloud international; mainland data under PRC rules | Alibaba publishes governance principles; IP complaint handling through Alibaba Cloud legal channels rather than a DMCA agent | Open models widely adopted; some public bodies apply the same device restrictions as to other PRC-hosted services |
| Kimi (Moonshot AI) | Beijing Moonshot Technology Co., Ltd. (China) PRC (Mainland) law; disputes in the courts of Beijing | Inputs, outputs and feedback licensed to Moonshot for model optimisation unless you contact the company to opt out | PRC | IP complaints by email or to the Beijing office, with identity and preliminary evidence required; no DMCA-style agent | Subject to the same government-device policies applied to PRC-hosted assistants in several countries |
Open-weight models (DeepSeek, Qwen and others) can be self-hosted, which removes the training and residency concerns listed for their hosted consumer services; the governing-law and support questions then fall on whoever hosts them.
What this means for a buyer
Five questions to settle before the model card
2. Are my people’s prompts training data?
Several consumer services train by default and bury the opt-out. Business tiers usually do not. Make sure staff are on the tier you think they are on.
3. Where does the data sit?
Storage location decides which governments can compel access. Residency options exist for most US and EU providers’ business tiers; PRC-hosted consumer services store in China.
4. Who answers a copyright notice?
A published DMCA-style procedure with a designated agent, counter-notice and repeat-infringer policy tells you the provider has thought about §512. Indemnity for outputs is now standard for US business tiers; ask for it.
5. Will my regulator or customer object?
Government device bans are a leading indicator of procurement trouble. If a public-sector customer cannot use your assistant, neither can you on their work.
How ELP GPT answers
Ontario law; no use of customer conversations to train an ELP model without separate notice and consent, with third-party processing disclosed; North American hosting with provider restrictions for business accounts; a published copyright procedure; no government restrictions. Where a competitor’s terms are better on a line, we will say so.
DMCA compliance
How ELP GPT handles copyright in practice
- Notice and takedown. A published procedure under 17 U.S.C. §512 and Canada’s notice-and-notice regime, a designated contact (copyright@elpgpt.com), counter-notice handling and a repeat-infringer policy. See the copyright policy.
- Output safeguards. The assistant is instructed not to reproduce substantial portions of copyrighted works, including lyrics and long passages, and to summarise and cite instead. Reports of reproduction are investigated and safeguards adjusted.
- Customer responsibility. Customers are responsible for the use of outputs and for the rights in what they submit, as set out in the Terms of Use.
- Enterprise indemnity. Enterprise agreements can include IP indemnity for outputs used within the terms and with safeguards enabled, on the same pattern as the US providers above.
A note on comparative claims
Comparative statements on this page are factual descriptions of published terms, made in the ordinary course of truthful comparative advertising. We do not disparage competitors; we describe what their documents say and link to sources. If a provider updates its terms, send the link and we will update the row with the date.
Sources
- Kimi Terms of Service (governing law, training licence, IP complaints)
- Alibaba Group AI governance
- Anthropic consumer data policy change, Sept 2025 (summary)
- DeepSeek government restrictions (overview, 2026)
- DeepSeek bans by countries and agencies
- OpenAI Copyright Shield announcement coverage
- Google generative-AI indemnification (law-firm briefing)
- Indemnification commitments across providers (WSGR)
Need the full legal pack?
Data-processing addendum, sub-processor list and provider map for your region.