Who Owns AI Output? Comparing ChatGPT, Claude, Gemini, and Grok

“You Own the Output” Can Mean Three Different Things

The four consumer contracts do not use one legal mechanism to describe AI output. OpenAI and Anthropic use an assignment: subject to the terms and applicable law, each assigns to the user whatever right, title, and interest the company may have in output. Google takes a different position. Its general Terms say Google will not claim ownership over original content generated by services that allow it. SpaceXAI says consumers retain ownership of User Content, a defined category that includes most Grok inputs and outputs, but follows that promise with an unusually broad license back to the company.

Those formulations can lead to a similar headline—“the user owns the output”—while allocating rights differently. An assignment attempts to transfer any provider-held rights. A non-claim says the provider will not assert ownership, but does not itself transfer a right. A license preserves one party’s ownership while authorizing another party to use the material.

The SpaceXAI license is the most expansive license-back in this four-provider comparison. It is described as irrevocable, perpetual, transferable, sublicensable, royalty-free, worldwide, and usable for any purpose. It reaches copying, modification, processing, publication, public display, derivative works, distribution, aggregation, product development, research, and other business purposes. If User Content contains a person’s image, likeness, voice, or similar attribute, SpaceXAI says the same rights apply to those attributes and makes the user responsible for obtaining the needed permissions.

Document evidence · SpaceXAI Consumer Terms · Our Use of User Content

Legally, that license is not the same thing as an assignment of ownership. Practically, however, it strips away much of what an ordinary person expects ownership to mean. If the provider may use the material for any purpose forever, transfer that permission, sublicense it, modify it, publish it, and build products from it without paying the user, the user retains title without retaining exclusive control. The phrase “you retain ownership” is therefore true in a narrow legal sense while potentially misleading as a description of the user's real leverage.

The rest of the contract makes that distinction more consequential. SpaceXAI pairs the license with broad consumer indemnity, a low general liability cap, a class-action waiver, a restrictive Texas forum, and fallback confidential arbitration. The company receives extensive rights to use the material while the consumer accepts limited recovery and a narrower path for challenging harm. Read together, those clauses matter more than the reassuring ownership sentence standing alone.

Google also takes a license over user content protected by intellectual-property rights. Its wording is worldwide, nonexclusive, royalty-free, and sublicensable, and permits hosting, using, distributing, communicating, modifying, and creating derivative works. The important difference is purpose. Google describes the license as limited to operating and improving services, promoting content the user shares publicly, and developing new technologies and services. Its duration is tied to the period during which the content is protected by intellectual-property rights, subject to removal and other exceptions.

OpenAI and Anthropic do not need a comparable output license to make their ownership promise work. Their terms instead reserve rights to process user content for service operation, safety, and improvement. That is why output ownership, content licensing, and training controls must be read separately.

Contract Ownership Does Not Create Copyright

The strongest consumer-facing ownership clause cannot transfer a copyright that never came into existence. The U.S. Copyright Office’s 2025 report states that copyright protects AI-assisted work when a human has determined sufficient expressive elements. It also concludes that prompting alone does not provide enough human control over the resulting expression. Human-authored selection, arrangement, or modification may still be protected.

This distinction changes what the contract comparison can establish. OpenAI and Anthropic assign any provider-held rights “if any.” That qualification matters. If an output contains no protectable human authorship, an assignment may leave the user with contractual permission to use the output but no exclusive copyright against others. Google’s promise not to claim ownership leads to the same practical limit. SpaceXAI’s recognition of user ownership is expressly subject to applicable law.

Inputs create a second limit. A provider cannot give a user ownership of third-party material merely because a model reproduced or transformed it. Each company places responsibility on users to have appropriate rights in what they submit and warns that output can be inaccurate, nonunique, or implicate third-party rights. None of the four consumer contracts is a warranty that every generated image, paragraph, recording, or code block is clear for commercial use.

The defensible answer is therefore narrower than most product copy: the contracts allocate whatever rights the providers have, but copyright law and third-party rights determine what actually exists to allocate.

Training Rights Are Separate From Output Ownership

A person can own an input or output while a provider remains authorized to process it. OpenAI’s Terms allow Content to be used to provide, maintain, develop, and improve services, while its Data Controls provide an opt-out for model improvement. The current help record says signed-in and signed-out users can opt out. Temporary Chats are not used for training, are deleted after 30 days, and may still be reviewed to monitor abuse.

Anthropic follows a similar consumer pattern. Its Terms allow Materials—inputs, outputs, and related interactions—to be used for service and product development, including model training, unless the user opts out in account settings. Feedback and material flagged for safety review remain stated exceptions. Anthropic’s help record says consumer users can choose whether chats and coding sessions improve Claude.

Google routes the decision through Gemini Apps Activity. Its privacy notice says chats associated with activity can be used to improve Google services, including generative models, and may be reviewed by humans. Chats disconnected from an account after human review may be retained for up to three years. With activity off, or when Temporary Chat is used, future chats are generally retained for 72 hours and are not used to train generative models unless the user submits feedback; safety and service-operation uses remain.

SpaceXAI gives logged-in consumers a choice about whether User Content is used for product development and model training. Its logged-out rule is different: where anonymous use is available and permitted, the Terms grant the company full rights to use data supplied to or obtained from the service for product development and training. That is not the same as the broader perpetual license, which appears earlier and is not written as though it disappears when training is disabled.

These controls answer a data-use question, not an ownership question. Turning training off can limit one purpose without retracting every license, retention exception, security use, legal hold, or feedback permission in the governing documents.

Every Provider Draws a Competitive Boundary Around AI Reuse

The most important market similarity is not in the ownership clause. It is in what users may do with the output afterward.

OpenAI prohibits using output to develop models that compete with OpenAI. Google’s general Terms prohibit using AI-generated content from its services to develop machine-learning models or related AI technology. Anthropic’s Consumer Terms go beyond a model-only restriction: they prohibit developing products or services that compete with Anthropic’s Services, including developing or training AI or machine-learning models and reselling the Services.

Document evidence · Google Terms of Service · Don’t abuse our services

SpaceXAI’s AUP is broader still. It prohibits using the service or any output to develop machine-learning models or any product or service that competes with SpaceXAI, directly or indirectly. Separate bullets prohibit scraping, harvesting, reselling, and distilling model data or outputs. The captured policy does not define “indirectly” or publish a limiting test for adjacent software, evaluation products, synthetic-data pipelines, or tools that assist another developer.

Document evidence · SpaceXAI Acceptable Use Policy · Competition Restriction

This is not a finding that every clause is enforceable in every setting. It is a finding about the contractual market boundary the providers are attempting to set. All four sell broad access to systems built from enormous accumulations of data and computation, yet all four restrict customers from turning the resulting output into some form of competing AI capability.

That asymmetry matters because output is not merely the last product in an AI supply chain. It can also become evaluation data, synthetic training data, retrieval material, model feedback, and scaffolding for new products. A provider that controls a widely used interface can gather interactions at scale while contractually restricting downstream firms from using the same interface to reduce their own dependence on the incumbent.

The Federal Trade Commission identified data, talent, and computational resources as essential generative-AI inputs. It warned that incumbents with large stores of user data may create barriers to entry and that API terms can be used to protect incumbency. The agency did not adjudicate these four contracts, and its 2023 analysis predates the versions reviewed here. It supplies the market mechanism: restrictions on data and output reuse can matter beyond an individual account because they shape who can accumulate the inputs needed to compete.

There is a legitimate provider interest on the other side. A company can protect infrastructure from scraping, prevent simple resale of a paid service, enforce safety controls, and safeguard proprietary model behavior. The difficult question is where protection of a service ends and foreclosure of adjacent innovation begins. The contracts do not answer that policy question; they stake out positions favorable to the provider.

The Market Fight Is Already Moving Beyond the Ownership Clause

The contracts describe what providers say users may do. Three current events show why those promises also have to be read against government policy, research credit, and control over model access.

On September 1, the U.S. Department of Justice filed a 20-page statement of interest in the consolidated OpenAI copyright litigation. It argues that using written works to train large language models is “exceedingly transformative” and that it would be legally incorrect to impose broad copyright liability that generally makes training impermissible without licensing. The filing also argues that mandatory licensing could leave only the largest technology companies able to pay, entrenching an oligopoly in model training.

That is the federal government’s advocacy position in pending litigation, not a court ruling. The filing also draws a boundary that matters for consumers and creators: acquisition, training, and output are separate uses. It argues that training may be fair while acknowledging that outputs which reconstruct and distribute protected expression can present different copyright questions. In other words, a favorable legal theory for training does not automatically clear a specific output for publication, sale, or ownership.

Document evidence · U.S. Department of Justice · Statement of Interest · Page 4

Above the Law’s Joe Patrice broadly agrees that training lawfully acquired material can be fair use when a model does not reproduce a market substitute, but criticizes the government’s intervention on institutional-conflict grounds. His September 3 analysis points to separate reporting that the federal government had discussed taking an equity stake in OpenAI, a potential interest the DOJ filing does not mention. That criticism does not prove the filing’s legal analysis wrong, nor does it establish that an equity agreement was completed. It does show why readers should examine who benefits from an official position as well as what the legal argument says.

A separate September 8 Axios report shows how the ownership issue changes when the input is unpublished research rather than public text. OpenAI announced that an internal model had produced a proof concerning the three-dimensional Navier–Stokes equations after its researchers heard that other mathematicians were close to major results. Axios reported that NYU mathematician Tristan Buckmaster questioned whether OpenAI had followed a research direction learned from his and Levent Alpöge’s work and raised concerns about private Codex material. OpenAI said it had not accessed their specific user data, but acknowledged it could not entirely rule out an indirect connection through de-identified product-use data that helped improve its models.

Those are contested allegations, not a finding that OpenAI took the researchers’ work or violated its terms. The episode still exposes a practical limit of contractual ownership. A researcher may retain formal rights in notes and outputs while lacking visibility into whether aggregated usage signals helped improve a provider that can later deploy far more compute against the same problem. Credit, confidentiality, training choice, and ownership become separate questions even when the contract uses reassuring assignment language.

Model access is another form of market power. On August 28, OpenAI announced that it intended to stop providing its models to Cursor on November 12 after SpaceX acquired the coding company. OpenAI cited a change-of-control provision in its custom agreement with Cursor and said it could not be confident SpaceX would comply with its terms. That announcement is OpenAI’s account of a commercial dispute, not an independent adjudication, and it concerns a negotiated supplier contract rather than the consumer Terms scored in this report. But it demonstrates a real downstream consequence: a provider can use contract rights to withdraw model access after ownership changes, affecting developers who selected a product partly because those models were available.

Read together, the asymmetry is sharper than an ownership sentence suggests. AI companies are arguing for broad legal freedom to learn from existing works, retaining contractual authority to use at least some customer interactions for improvement, restricting customers from using output to build competing AI, and policing access to their models through private agreements. Those positions are not legally identical and should not be collapsed into one accusation. They nevertheless shape the same market question: who gets to turn other people’s material, activity, and dependency into durable competitive advantage?

Liability and Dispute Terms Determine the Value of the Promise

Rights over output matter most when something goes wrong: an output infringes, an agent takes an unwanted action, a confidential input is retained, or a business builds on material it cannot safely use. The four contracts transfer that risk through liability limits, indemnity provisions, and dispute procedures.

Anthropic and SpaceXAI impose broad indemnity obligations on ordinary consumers. In general terms, indemnity asks the user to protect the provider and covered parties from certain third-party claims and costs arising from the user’s content, use, products, misconduct, or violation of the terms. OpenAI’s and Google’s captured U.S. terms place comparable indemnity language in their business-or-organization provisions rather than making it a general consumer obligation.

Anthropic generally caps its aggregate liability at the greater of $100 or the amount paid during the preceding six months. OpenAI uses the greater of $100 or the amount paid in the preceding 12 months. Google’s current captured U.S. Terms use the greater of $200 or 12 months of fees. SpaceXAI uses the greater of $100 or the amount paid under the contract. Each clause contains qualifications, exceptions, and acknowledgments that local law may preserve additional rights.

Dispute routes also differ. OpenAI requires individual arbitration and a class-action waiver but provides a 30-day opt-out process. SpaceXAI’s captured Terms waive class participation, select courts in Wichita or Tarrant County, Texas, and use confidential individual arbitration if the forum-selection clause is unenforceable. Its class waiver expressly extends to claims against X, Cursor, and SpaceX entities.

Anthropic’s captured Consumer Terms use California law and courts in San Francisco without a comparable arbitration or class-waiver provision. Google’s captured U.S. Terms use California law and courts in Santa Clara County and likewise contain no arbitration or class-action waiver in the general terms reviewed.

These differences do not decide which service is safest. A liability cap may be limited by consumer law; an arbitration clause may not govern every dispute; and a provider may offer additional protections under a business or enterprise contract. They do show why “you own your output” is incomplete as a purchasing claim. The economic value of a right depends partly on who bears the loss and where the user can seek a remedy.

Which AI Has the Best Consumer Legal Terms?

On the captured documents, OpenAI ranks first, Google and Anthropic tie for second, and SpaceXAI ranks fourth. This is a ranking of consumer-facing legal terms—not answer quality, price, safety performance, privacy engineering, or enterprise protections.

The rubric gives each provider zero, one, or two points on five equally weighted questions: How much practical control does the user retain over output? How restrained are the provider's retained content-use rights? Can a consumer decline model training? How much freedom remains to use output in downstream AI development? How balanced are the liability, indemnity, arbitration, and class-action provisions? The maximum is 10 points.

OpenAI scores 7. Its affirmative assignment, training opt-out, Temporary Chat treatment, comparatively narrow competing-model restriction, and lack of a general consumer indemnity make it the strongest overall package in this corpus. Its $100-or-fees liability cap, individual arbitration, and class waiver keep it from scoring higher, although the arbitration provision includes a 30-day opt-out.

Google and Anthropic each score 6 for different reasons. Google has the most favorable captured dispute posture: no general consumer indemnity, arbitration clause, or class waiver was found in the U.S. Terms reviewed, and its liability cap uses $200 rather than $100 as the fixed-dollar comparison. But it makes a non-claim instead of assigning output rights, retains a purpose-based content license, and restricts using generated content to develop models or related AI technology. Anthropic gives users a clearer output assignment and a strong training choice, but its ban reaches competing products or services, and its Terms combine consumer indemnity with a $100-or-six-months liability cap.

SpaceXAI scores 1. Its logged-in training choice earns the only point. The ownership label earns none on practical control because the perpetual, transferable, sublicensable, any-purpose license leaves the consumer without meaningful exclusivity. The different logged-out training rule, broad direct-or-indirect competition restriction, consumer indemnity, low liability cap, class waiver, and Texas or confidential-arbitration dispute structure complete the weakest package in this corpus.

The score is deliberately reproducible rather than definitive. Different weights change the outcome. A person who cares most about avoiding arbitration may prefer Google. A creator who values an affirmative assignment and training control may favor OpenAI or Anthropic. A company buying enterprise protections should not use this consumer ranking at all; it should compare the actual business contracts offered for its account.

What These Terms Likely Mean for the Future of Work

The first change is that AI fluency will include contract fluency. Workers will not only need to know how to produce a useful draft, image, analysis, or software component. They will need to know whether the chosen service may train on the underlying material, what license the provider retains, whether another person's likeness or confidential information can be submitted, and whether the result may enter a competing product. Researchers and other knowledge workers will also need to distinguish formal ownership from the confidentiality, attribution, and training rules that determine whether an unpublished idea remains meaningfully under their control. That turns model selection from a personal preference into part of ordinary information governance.

The second change is that proof of human contribution will become more valuable. Contract language can allocate provider-held rights, but it cannot manufacture copyright. For commercially important work, a prompt-and-download workflow leaves a weak record of authorship. Version history, source files, edits, selection decisions, arrangement, and documented human judgment can help show what the person contributed. The valuable worker is less likely to be the person who merely elicited an output than the person who can direct, verify, transform, and account for it.

The third change is that employers will procure legal risk, not just model capability. A public chatbot may be adequate for casual ideation and still be the wrong instrument for client files, unreleased product plans, regulated decisions, or autonomous actions. Organizations will increasingly separate low-risk consumer use from approved business accounts with negotiated retention, confidentiality, indemnity, security, and intellectual-property terms. The model leaderboard will matter less when the contract assigns an unacceptable exposure to the employer.

The fourth change is a likely widening gap between people who can use AI and firms that can build on AI. All four providers restrict some downstream AI development. As dominant interfaces accumulate user interactions and feedback, smaller builders may be allowed to consume outputs for ordinary tasks while being prevented from using them to train, evaluate, distill, or improve a competing system. That could push innovation toward open-weight models, licensed datasets, first-party data, and vendors whose contracts expressly permit the intended development work.

Agentic systems add a fifth pressure: authority can expand faster than accountability. When a tool can browse, send messages, change files, execute code, or interact with accounts, its output is no longer only text. It is an action with financial, reputational, privacy, and operational consequences. If the provider disclaims those consequences while the worker or employer remains responsible, organizations will need approval gates, spending limits, audit logs, scoped credentials, and human review around consequential actions.

These are projections from the contractual incentives in the reviewed documents, not measured forecasts of employment. They identify the skills and controls the contracts reward: provenance, judgment, permission management, vendor evaluation, human authorship, and accountable oversight.

How to Make an Educated AI Choice

Start with the work, not the brand. A defensible decision can be made in seven steps:

1. Classify the material. Separate casual public information from confidential business data, client records, regulated data, unreleased creative work, and uploads containing another person's face or voice.
2. Classify the intended output use. Ordinary drafting, public publication, client delivery, software integration, model evaluation, synthetic-data creation, and training another system create different contractual risks.
3. Read the complete governing package. Check the consumer or business terms, incorporated use policy, privacy notice, training controls, retention rules, and any product-specific or agentic terms. The ownership sentence is not enough.
4. Choose the correct account mode. Use available training controls and temporary modes before submitting material. Do not assume logged-out use is more private; the SpaceXAI documents reviewed show why account mode can change the rights granted.
5. Preserve provenance. Save the effective terms, relevant settings, prompts, source material, intermediate versions, and substantial human edits for work whose ownership or accuracy may matter later.
6. Limit autonomous authority. Give agents the fewest permissions, funds, files, and external accounts necessary. Require human approval before publication, payment, deletion, legal communication, or another difficult-to-reverse action.
7. Escalate when the downside is material. Consumer terms are a poor procurement shortcut for sensitive or high-value work. Seek business terms and qualified legal review when confidentiality, copyright, indemnity, retention, regulated decisions, or competitive AI development could affect the organization.

No provider is the best choice for every task. Based on these documents, OpenAI offers the strongest general consumer package, Google offers the most favorable captured dispute posture, Anthropic pairs strong output and training language with stricter competitive and indemnity terms, and SpaceXAI asks consumers to accept the broadest retained rights and risk transfer. The educated choice is the provider-account-use combination whose contract matches the actual work.

Methodology and Limitations

This comparison reviewed 14 materially used records as of September 10, 2026. The core 10-source corpus comprised U.S.-facing consumer terms for Anthropic, OpenAI, Google, and SpaceXAI; the SpaceXAI Acceptable Use Policy incorporated into its Consumer Terms; first-party training and privacy records; U.S. Copyright Office guidance; and Federal Trade Commission competition analysis. The current-events extension reviewed the full 20-page DOJ statement of interest, full reporting from Above the Law and Axios, and OpenAI’s first-party Cursor announcement.

Each provider was coded on the same dimensions: output-rights mechanism, provider license or content-use authority, consumer training choice, AI-development restriction, liability cap, consumer indemnity, arbitration, and class-waiver language. “Broader” in the competitive-use discussion refers only to the literal categories named in the captured contract; it is not a legal severity score or a prediction about enforcement.

The consumer-terms ranking is an author-calculated comparison derived from that coding. Each provider received zero, one, or two points on practical output control, provider content-use restraint, consumer training choice, downstream AI freedom, and consumer remedy balance. The five dimensions were equally weighted. Tied totals remain tied. Practical output control treats formal ownership as weaker when a provider simultaneously receives permanent, transferable, sublicensable, any-purpose rights that remove meaningful exclusivity. The rubric does not measure model quality, price, security controls, actual privacy performance, enforceability, regional law, or enterprise-contract protections.

The corpus is version-sensitive. The OpenAI, SpaceXAI Terms, and SpaceXAI AUP pages captured for this report label the displayed text a previous version. Google’s U.S. Terms page shows an effective date of July 30, 2026, superseding the May 22, 2024 date in the supplied research note. Anthropic’s extracted Consumer Terms did not expose a visible effective date. Regional terms, service-specific terms, paid-business contracts, and product interfaces may differ. The three current events are contextual case studies rather than additional consumer-contract rows, so they do not change the equal-weight provider ranking.

This is comparative contract analysis, not legal advice. It does not determine whether an output is copyrightable, whether a training use infringes, whether a particular downstream product competes with a provider, or whether any clause would be enforced in a specific jurisdiction.

Conclusion: Output Ownership Is a Market Rule, Not Just a User Promise

The four providers all give consumers a path to use generated output, but they do not make the same legal promise. OpenAI and Anthropic assign provider-held output rights. Google declines to claim ownership. SpaceXAI recognizes user ownership while taking a permanent, transferable license back.

None of those promises guarantees copyright. All four providers also draw a line around using output to build AI, and that line ranges from competing models to related AI technology, competing services, and direct or indirect competition. The DOJ’s new filing, the Navier–Stokes credit dispute, and OpenAI’s decision to wind down model supply to Cursor show different parts of the same market structure: the law governing training, the opacity surrounding product-use data and credit, and the private contracts governing access can matter as much as formal title to an output.

The useful question is therefore not simply “Who owns AI output?” It is: what right exists, what did the provider keep, what uses are forbidden, and who carries the risk when the answer is wrong?

Sources

1. Anthropic. (n.d.). Consumer terms. https://www.anthropic.com/legal/consumer-terms
2. Anthropic. (2026, March 16). Is my data used for model training? https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training
3. OpenAI. (2026, January 1). Terms of use (captured page labels this a previous version). https://openai.com/policies/terms-of-use/
4. OpenAI. (n.d.). Data controls FAQ. https://help.openai.com/en/articles/7730893-data-controls-faq
5. Google. (2026, July 30). Google terms of service—United States. https://policies.google.com/terms
6. Google. (2026, August 10). Gemini Apps privacy hub. https://support.google.com/gemini/answer/13594961
7. SpaceXAI. (2026, September 1). Terms of service—Consumer (captured page labels this a previous version). https://x.ai/legal/terms-of-service
8. SpaceXAI. (2026, August 14). Acceptable use policy (captured page labels this a previous version). https://x.ai/legal/acceptable-use-policy
9. U.S. Copyright Office. (2025, January 29). Copyright Office releases part 2 of artificial intelligence report. https://www.copyright.gov/newsnet/2025/1060.html
10. Federal Trade Commission. (2023, June 29). Generative AI raises competition concerns. https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2023/06/generative-ai-raises-competition-concerns
11. United States Department of Justice. (2026, September 1). Statement of interest of the United States, In re OpenAI, Inc. copyright infringement litigation, No. 25-md-3143, Document 316. https://abovethelaw.com/wp-content/uploads/sites/4/2026/09/AI-Brief.pdf
12. Patrice, J. (2026, September 3). DOJ tells court AI training is fair use, forgets to mention it’s negotiating a stake in OpenAI. Above the Law. https://abovethelaw.com/2026/09/doj-tells-court-ai-training-is-fair-use-forgets-to-mention-its-negotiating-a-stake-in-openai/
13. OpenAI. (2026, August 28). Our decision on Cursor following its acquisition by SpaceX. https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/
14. Mills, M. (2026, September 8). OpenAI’s historic math solution overshadowed by credit controversy. Axios. https://www.axios.com/2026/09/08/openai-math-solution-navier-stokes-credit