Two-Tier Privacy: Consumer vs. Enterprise Terms

What this page is: A comparison of every GenGA provider that publishes both a consumer-facing Terms of Service and a separate enterprise/business contract (Commercial Terms), testing whether the same company offers categorically different data-rights protection depending on whether the customer pays for an enterprise tier.

Providers With Confirmed Dual-Tier Documents

Four of the 11 GenGA providers have both a consumer Terms of Service and a Commercial Terms document captured in this dataset: ChatGPT (OpenAI), Claude.ai (Anthropic), Le Chat (Mistral), and xAI. The other 7 providers either have no captured enterprise-tier document (Meta AI, Qwen Chat), only an API/developer-tier document without a clear consumer/enterprise split in the same product line (DeepSeek’s Developer Terms vs. Terms of Service is a narrower API-vs-consumer split, not a full enterprise contract; Llama API and Perplexity’s Developer Terms are similarly API-tier-only without a comparable “Commercial Terms” enterprise contract), or are out of scope for this comparison (Microsoft Copilot, Google Generative AI Services).

The Comparison

Provider	Document Type	Input License Scope	Output Restriction Scope	Key Difference
OpenAI	ChatGPT Terms of Service (consumer)	Default opt-out training; broad worldwide right to use Content “to provide, maintain, develop, and improve our Services”	Bans using Output to develop competing AI models	Training is the default; the user must act to stop it
OpenAI	ChatGPT Commercial Terms	“Will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees” (§4.2) — training requires affirmative opt-in	Same competing-model ban, with a narrow Permitted Exception for non-distributed classifiers	Training flips from opt-out to opt-in; OpenAI additionally confirms it gets no IP rights in Customer Content (§9.1)
Anthropic	Claude.ai Terms of Service (consumer)	Training-eligible by default unless the user opts out via account settings; two carve-outs survive the opt-out regardless (Feedback-rated conversations, safety-flagged content)	Bans developing/training “any artificial intelligence or machine learning algorithms or models” that compete with Anthropic	Opt-out exists but is incomplete — certain content trains regardless of the user’s choice
Anthropic	Claude.ai Commercial Terms	“Anthropic may not train models on Customer Content from Services” — unconditional, no opt-in exception of any kind described in the document	Same competing-model ban, phrased as a Use Restriction	The strongest no-training commitment of any document in this entire wiki — no carve-outs survive, unlike the consumer tier
Mistral	Le Chat Terms of Service (consumer)	Training-eligible by default unless on Le Chat Enterprise/paid API tier (or, depending on capture date, Le Chat Pro/Student), subject to opt-out	Narrow image-Output-only competing-model ban	The consumer product is the one that trains by default; only the paying/enterprise tiers are protected
Mistral	Le Chat Commercial Terms	Eroded over time, the only reversal in either direction found in this comparison: 2025-11-11 baseline had an unconditional no-training commitment (“we do not have access to the Infrastructure or Your Data and do not use Your Data to train our Models”); by 2025-12-15 this became a conditional opt-out structure, then progressively broadened (2026-03-12 added a Le Chat Pro/free-tier training-eligible-by-default carve-out; 2026-04-08 extended it to Le Chat Teams)	Same narrow image-Output ban; a broader web-search-Output restriction existed transiently (2025-12-15) and was removed by 2026-03-12	Mistral is the only provider in this comparison where the enterprise tier’s protection degraded toward the consumer tier’s weaker stance, rather than the consumer tier staying weak while enterprise stayed strong
xAI	Terms of Service (consumer)	Maximal and unconditional: “an irrevocable, perpetual, transferable, sublicensable, royalty-free, and worldwide right… for any purpose,” not gated by the separate training opt-out; logged-out users get no opt-out at all	Bans using Output to develop competing models; mandatory (not merely recommended) AI-disclosure duty	The broadest Input license of any document examined anywhere in this comparison, paired with a tiered opt-out that fails entirely for anonymous users
xAI	Commercial Terms	Purpose-limited license only (“to provide the Services… and to enforce xAI policies, prevent abuse, and perform safety/compliance/moderation” — no “improve services” catch-all) plus a near-unconditional no-training commitment: “xAI shall not use any User Content for any of its internal AI or other training purposes,” with one carve-out for de-identified data	Bans representing Output as human-generated or using Output to train the customer’s own competing models	The gap between xAI’s two tiers is the largest of any provider in this comparison — from the dataset’s single broadest consumer license to one of its four unconditional enterprise no-training guarantees

The Gap, Quantified

Three of the four providers (OpenAI, Anthropic, xAI) show the predicted “protection as a paid feature” pattern cleanly: the enterprise/Commercial Terms document either requires affirmative opt-in for training (OpenAI) or forecloses training entirely (Anthropic, xAI), while the consumer Terms of Service trains by default, with the user bearing the burden of opting out — and in xAI’s case, no opt-out at all for logged-out use. Anthropic’s and xAI’s Commercial Terms are 2 of the 4 unconditional no-training commitments found anywhere in the GenGA dataset (the other 2 are Llama API’s and Perplexity’s Developer Terms — both API-tier documents, reinforcing that unconditional protection in this dataset is consistently a paid/professional-tier feature, never a consumer-tier default).

Mistral is the instructive exception, not a counter-example. Its enterprise tier did not start weak — it started as the strongest commitment in the comparison (“we do not have access to the Infrastructure or Your Data”) — but eroded toward the consumer tier’s permissiveness within about four months of captures. This suggests the bifurcation pattern is not a fixed structural feature of “enterprise contracts are inherently safer,” but a current snapshot that a provider can move in either direction; Mistral’s trajectory is the dataset’s only documented case of the gap closing by the enterprise side getting worse, rather than the consumer side getting better.

Research Significance

This pattern complicates the “Data Colonialism 2.0” framing in a specific way: the asymmetry isn’t simply “AI providers vs. users,” it’s “AI providers’ paying enterprise customers vs. their ordinary consumer users,” with the company’s own contractual posture toward AI training varying by who is paying for access, not by any consistent policy on training itself. Pandit et al. (2026)‘s finding that GenAI terms place “responsibilities [users] cannot materially fulfil” onto consumers takes on a sharper edge here: an enterprise customer can negotiate and obtain an unconditional no-training guarantee, while a consumer is offered, at best, an opt-out with carve-outs (Anthropic, Le Chat) or, at worst, no opt-out at all for a category of use (xAI’s logged-out users). Atkinson (2025)‘s contractual-asymmetry framework — built around the question of who has genuine notice and bargaining power when a contract term is imposed — applies directly: enterprise customers negotiate or at least select Commercial Terms with legal review, while consumer users encounter the same kind of broad, take-it-or-leave-it Terms of Service that Atkinson’s paper treats as enforceable against third parties precisely because consent is presumed from access, not from negotiation.

Limitations

This comparison covers only 4 of 11 GenGA providers — the 7 without a captured dual-tier document are not evidence either way; they may have an enterprise tier this wiki’s source captures simply did not include (Open Terms Archive’s genai-eu export determines what was captured, not this wiki).
“Input License Scope” and “Output Restriction Scope” summaries above are this wiki’s own LLM-assigned characterizations of each document’s Overview section, not a verbatim legal comparison; see each linked platform page for exact clause text.
The Mistral/Le Chat trajectory is the only one in this comparison with enough dated captures to show change over time within a single tier; for the other three providers, this page reflects a snapshot, not a confirmed trend.