Redistributed

Documents: OpenAI Terms of Use, Services Agreement, Usage Policies Source: OpenAI Policies

OpenAI's terms govern what over 300 million people can do with ChatGPT. They determine who owns AI-generated content, who bears the risk when things go wrong, and what happens to every conversation users have with the system. Reading them carefully reveals an asymmetry common to platform businesses: users provide value and bear risk; the platform captures value and disclaims risk. This is not unique to OpenAI — but the scale of AI deployment makes it worth examining carefully.

The Two-Track System

OpenAI maintains separate terms for consumers and businesses — a split that reveals who they consider valuable.

For individual users (Terms of Use):

Your conversations may train future models (unless you opt out)
You own your outputs, but so might other users who get the same response
OpenAI can change terms at any time
Maximum liability: $100 or what you paid in the last 12 months

For business users (Services Agreement, updated January 2026):

Data is never used for training by default
Explicit ownership assignment: "OpenAI hereby assigns to Customer all OpenAI's right, title, and interest, if any, in and to Output"
IP indemnification for Enterprise customers
More structured change processes

The two-tier structure is familiar from other platform businesses, but the stakes are different with AI. Consumer users are contributing something beyond engagement — they are providing training signal that directly improves the product's core capability. For context on how this fits the broader platform dynamic, see API Economy.

Data Usage: The Training Question

OpenAI's Help Center states:

"When you use services such as ChatGPT, Sora, or Operator, we may use your content to train our models."

This language accomplishes two things. First, it normalizes the extraction of value from user interactions — every prompt teaches the model, every conversation improves the product. Second, it makes opting out the user's burden.

The opt-out process requires navigating to OpenAI's privacy portal and clicking "do not train on my content." Most users never find this. Those who do must weigh whether disabling training also disables chat history — a dark pattern that conflates convenience with consent.

For API and Enterprise customers, the terms are different:

"OpenAI will only use Customer Content as necessary to provide Customer with the Services, comply with applicable law, enforce the OpenAI Policies, and prevent abuse. OpenAI will not use Customer Content to develop or improve the Services."

This two-tier system creates a redistribution mechanism worth naming: casual users contribute training data for free; paying customers get protection. The labor of ordinary users subsidizes the product that enterprises pay for. The Hidden Labor analysis examines parallel dynamics in other parts of the AI supply chain.

Output Ownership: A Gift That Keeps Giving (to Others)

The terms assign output ownership to users. Sounds generous. But read the fine print:

"Due to the nature of our Services and artificial intelligence generally, output may not be unique and other users may receive similar output from the Services."

And:

"OpenAI's assignment does not extend to other users' output or any Third Party Output."

So you own your output — but someone else might get the exact same output, and they own that too. Both of you have valid ownership claims to identical content. Neither can exclude the other.

This isn't ownership in any traditional sense. It's more like owning a copy of a book: you have rights to your copy, but so does everyone else who bought one. The difference is that with books, scarcity creates value. With AI outputs, abundance destroys it.

The practical implication: Don't build a business on the assumption that your ChatGPT outputs are exclusive. They're not.

Liability: You're On Your Own

The warranty disclaimer could hardly be more comprehensive:

"THE SERVICES ARE PROVIDED 'AS IS.' OPENAI AND ITS AFFILIATES AND LICENSORS MAKE NO WARRANTIES (EXPRESS, IMPLIED, STATUTORY OR OTHERWISE) WITH RESPECT TO THE SERVICES, AND DISCLAIM ALL WARRANTIES INCLUDING WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, SATISFACTORY QUALITY, NON-INFRINGEMENT, AND QUIET ENJOYMENT."

And:

"OPENAI DOES NOT WARRANT THAT THE SERVICES WILL BE UNINTERRUPTED, ACCURATE OR ERROR FREE, OR THAT ANY CONTENT WILL BE SECURE OR NOT LOST OR ALTERED."

The limitation of liability follows:

"OPENAI'S AGGREGATE LIABILITY UNDER THESE TERMS WILL NOT EXCEED THE GREATER OF THE AMOUNT YOU PAID FOR THE SERVICE THAT GAVE RISE TO THE CLAIM DURING THE 12 MONTHS BEFORE THE LIABILITY AROSE OR ONE HUNDRED DOLLARS ($100)."

A free ChatGPT user who relies on medical advice from the model and is harmed has recourse to exactly $100. A Plus subscriber paying $20/month could recover at most $240.

The terms explicitly warn:

"You accept and agree that any use of outputs from our Service is at your sole risk and you will not rely on output as a sole source of truth or factual information, or as a substitute for professional advice."

In practice: if ChatGPT hallucinates a legal case that gets you sanctioned, if it gives medical advice that harms you, if it provides code that introduces security vulnerabilities — the terms place that risk on you. OpenAI built the system, trained it, deployed it, and profits from it. The risk allocation runs one way. This pattern is standard in software licensing, but as AI systems are deployed in higher-stakes contexts — healthcare, legal research, financial analysis — the gap between capability marketing and liability acceptance becomes harder to justify. The AI in Healthcare explainer examines what this looks like in practice.

Prohibited Uses: The Evolving Redlines

OpenAI's Usage Policies define what you cannot do. The current list includes:

Absolute prohibitions:

Child sexual abuse material (CSAM)
Weapons of mass destruction (CBRN weapons)
Terrorism
Human trafficking and forced labor

Restricted without oversight:

Medical or legal advice without professional involvement
High-stakes automated decisions without human review
Political campaigning at scale

Recently clarified:

Malicious cyber activity
Covert influence operations
Surveillance for authoritarian regimes

But the history of these policies reveals something important: they change.

The military policy reversal: In January 2024, OpenAI quietly removed language explicitly banning use of its technology for "military and warfare." Until January 10, 2024, the usage policy included a ban on "activity that has high risk of physical harm, including... weapons development... military and warfare."

After the change, OpenAI replaced this with a broader prohibition on using services "to harm yourself or others." An OpenAI spokesperson told The Intercept that the goal was "to provide clarity and allow for military use cases the company does agree with."

By June 2024, OpenAI announced a pilot with the Department of Defense's Chief Digital and Artificial Intelligence Office. The ceiling: $200 million.

What changed? Not the technology's capabilities. Not new safety research showing military AI is safe. The business opportunity changed. The terms followed. This pattern — where usage policies track commercial incentives rather than safety research — is worth watching across the industry. See Regulatory Capture for how business interests shape governance more broadly.

Terms Can Change at Will

The services agreement grants OpenAI broad authority to modify terms:

Users' "continued use" of the services after changes constitutes acceptance. The alternative is to stop using the service entirely.

This creates a ratchet. Users invest in the platform — build workflows, train teams, develop dependencies. Then the terms change. Users can accept the new terms or abandon their investment. Most accept.

August 2025: The privacy reversal

In August 2025, all three major AI providers — OpenAI, Google, and Anthropic — weakened user privacy protections:

August 26: OpenAI announced a monitoring system that scans all ChatGPT conversations for potentially harmful content, with human reviewers authorized to read chats and report users to law enforcement
August 28: Anthropic made data collection the default setting for all users
September 2: Google made all chats contribute to training by default unless users opt out

The industry moved in lockstep. When every provider changes simultaneously, users have nowhere to go. The "choice" to accept new terms is no choice at all — a dynamic the AI and Antitrust analysis examines in depth. This is one reason why Running AI Locally, Ollama, and other self-hosted alternatives matter: they provide an exit option that disciplines platform behavior, at least for technically capable users.

The Monitoring Question

The August 2025 change deserves special attention. OpenAI's new system:

Scans all conversations for "potentially harmful content"
Escalates flagged conversations to human reviewers
Authorizes reviewers to read chats and report users to law enforcement

The definition of "potentially harmful" is OpenAI's to make. The threshold for escalation is opaque. The process for contesting a report is unclear.

This creates a surveillance infrastructure governed by a private company's judgment. Users accepted it by continuing to use ChatGPT after the terms changed. Most didn't notice.

Comparison: OpenAI vs. Competitors

How do OpenAI's terms compare to Anthropic and Google?

Aspect	OpenAI	Anthropic	Google
Consumer data training	Yes (opt-out)	Yes (opt-out after Aug 2025)	Yes (opt-out)
API data training	No (by default)	No (by default)	No (by default)
Output ownership	Assigned to user	Assigned to user	Assigned to user
IP indemnity	Enterprise only	Enterprise only	Some products
Military use	Allowed (since 2024)	Varies	Varies
Liability cap	$100 or 12 months fees	Similar	Similar

The convergence is notable. Competition on terms has collapsed. All major providers:

Train on consumer data by default
Protect enterprise data
Disclaim all warranties
Cap liability near zero
Reserve the right to change terms unilaterally

When everyone offers the same terms, there is no market pressure to improve them. This convergence is itself an argument for regulatory intervention — the EU AI Act and Who Regulates Foundation Models explore what that intervention might look like. It is also an argument for Public AI alternatives that operate under different incentive structures entirely.

The Redistribution Analysis

Read structurally, OpenAI's terms accomplish a specific redistribution:

Value flows up:

User conversations train models (for free users)
User feedback improves products
User content reveals preferences, use cases, market opportunities
The collective intelligence of millions of interactions becomes OpenAI's asset

Risk flows down:

Users bear all responsibility for outputs
Users must verify accuracy
Users face liability if outputs cause harm
Users have no recourse if the service fails

Control remains centralized:

OpenAI sets the rules
OpenAI changes the rules
OpenAI decides what content is acceptable
OpenAI determines who gets reported to law enforcement

This is extraction with extra steps. The "free" tier of ChatGPT is not free — it is paid in data, training signal, and risk assumption. The terms formalize this exchange while obscuring it. This pattern is common to platform economics (see Pricing Power), but AI intensifies it: the value users contribute is not just attention or content but direct improvement of the system's capabilities.

What to Do With This

If you're a casual user:

Opt out of training data if you care about privacy (Settings > Data Controls)
Don't share sensitive information — assume it's monitored
Never rely on ChatGPT for medical, legal, or financial decisions
Remember: $100 is the most you can recover if things go wrong

If you're a business:

Use API or Enterprise tiers to avoid training data extraction
Negotiate specific terms if your use case requires it
Build contingency plans for terms changes
Don't assume outputs are exclusive or protectable

If you're a policymaker:

Notice the convergence among providers on user-hostile terms — it signals a market failure
Consider whether "consent" via terms of service is meaningful when alternatives are absent
Examine whether liability caps are appropriate for AI systems deployed in high-stakes contexts
Evaluate whether training data extraction requires clearer consent
Look at how the EU AI Act handles similar questions, and consider what Democratic AI Governance could look like domestically

Llama License — Meta's different approach to terms and access
API Economy — The platform dynamics shaping AI terms
Who Regulates Foundation Models — The governance gap these terms exploit
AI and Antitrust — Why provider convergence on terms matters
Regulatory Capture — How business interests shape AI governance
Public AI — Alternatives to corporate platform terms
Running AI Locally — The exit option that open weights provide
Democratic AI Governance — What user-centered governance could look like
OpenAI For-Profit — The corporate structure behind these terms
Hidden Labor — The parallel value extraction in AI's supply chain

Sources

OpenAI. "Terms of Use." OpenAI Policies
OpenAI. "Services Agreement." Effective January 1, 2026. OpenAI Policies
OpenAI. "Usage Policies." OpenAI Policies
OpenAI. "How Your Data Is Used to Improve Model Performance." Help Center
The Intercept. "OpenAI Quietly Deletes Ban on Using ChatGPT for 'Military and Warfare'." January 12, 2024. Article
TV News Check. "OpenAI, Google & Anthropic All Just Quietly Backtracked User Privacy Settings." August 2025. Article
OpenAI. "Disrupting Malicious Uses of AI." October 2025. Report
Terms.Law. "Who Owns AI Content? ChatGPT, Claude, Midjourney & Gemini Rights Compared." 2026. Analysis

General

Doc Review: OpenAI's Terms of Service