Who Owns ChatGPT Outputs? OpenAI Terms, Copyright & Commercial Rights

OpenAI's Terms of Service contain a clear ownership assignment clause:

This appears straightforward, but the phrase "to the extent permitted by applicable law" is crucial. It means:

• Input ownership: You retain rights to your prompts and any content you provide
• Output assignment: OpenAI transfers whatever rights it may have in outputs to you
• Legal limitations: This doesn't override copyright law - if the output isn't copyrightable, there's nothing to own
• Non-exclusivity: Other users might generate identical or similar outputs

Free vs Plus vs API vs Enterprise

Feature	Free	Plus ($20/mo)	API	Enterprise
Output ownership	✓ Yes	✓ Yes	✓ Yes	✓ Yes
Training opt-out	✗ No	✓ Yes	✓ Yes (default)	✓ Yes (default)
Commercial use	✓ Allowed	✓ Allowed	✓ Allowed	✓ Allowed
IP indemnification	✗ No	✗ No	✗ Limited	✓ Yes
Data retention	30 days	30 days	0 days (API)	Custom

The legal landscape for AI-generated content has evolved dramatically since this analysis was first published. Here's what has changed—and what remains uncertain—as of early 2026.

1. NYT v. OpenAI: Case Proceeds to Discovery (March 2025)

In March 2025, Judge Sidney Stein of the Southern District of New York rejected OpenAI's motion to dismiss the New York Times' copyright infringement lawsuit. The court found that the complaint sufficiently alleged that ChatGPT could serve as a "market substitute" for reading actual NYT articles—the key question for copyright liability.

• Fair use in flux: OpenAI's primary defense—that training on copyrighted works is fair use—received mixed signals from the court. The judge did not rule on the fair use question at the motion to dismiss stage, meaning it will be decided later based on evidence developed during discovery.
• Output-side claims survive: The NYT's claim that ChatGPT sometimes reproduces near-verbatim excerpts of paywalled articles was deemed plausible enough to proceed.
• No trial date set: As of early 2026, the case is still in discovery. Legal experts do not expect a trial before late 2026 or 2027.

2. Discovery Battle: 20 Million ChatGPT Logs Ordered (January 2026)

In a landmark procedural ruling in January 2026, Judge Stein ordered OpenAI to produce 20 million anonymized ChatGPT conversation logs to plaintiffs in the consolidated copyright cases. This is the largest discovery request ever granted in AI copyright litigation.

• Privacy argument rejected: OpenAI argued that producing user conversations would violate user privacy. The court held that users who voluntarily submit communications to a chatbot operated by a for-profit company do not have a reasonable expectation of privacy in the content of their prompts.
• Anonymization required: OpenAI must anonymize personally identifiable information before production, but the content of the prompts and outputs must be disclosed.
• Multi-billion dollar question: Discovery will likely reveal how often ChatGPT reproduces copyrighted content, which could determine liability exposure across dozens of pending lawsuits.
• Consolidated cases: Over a dozen OpenAI copyright cases have been consolidated under Judge Stein in the Southern District of New York, including claims by authors, journalists, and media companies.

3. Output-Side Copyright Rulings: "Substantial Similarity" Test Applied

Courts have begun to distinguish between training-side infringement (using copyrighted works to train the model) and output-side infringement (whether the AI's responses infringe copyright). Two key rulings emerged in 2025:

Silverman v. OpenAI (July 2025)

The Ninth Circuit affirmed dismissal of Sarah Silverman's output-side copyright claim against OpenAI. The court held that ChatGPT's summaries of her books did not constitute "substantial similarity" under copyright law because they were transformative and did not reproduce protectable creative expression.

• Key holding: A factual summary or synopsis of a copyrighted work is not infringement if it does not reproduce the author's original creative expression.
• Training claims survive: Silverman's claim that OpenAI unlawfully trained on her books was allowed to proceed—training and output are separate legal questions.

Perplexity Case (October 2025)

In contrast, media plaintiffs suing Perplexity AI survived a motion to dismiss by alleging that Perplexity's AI search engine sometimes reproduces articles verbatim—not just summaries. This distinction may be outcome-determinative:

• Summaries and transformative outputs = likely not infringement
• Verbatim or near-verbatim reproduction = potential infringement

4. Supreme Court Declines Thaler v. Perlmutter: No Copyright for Pure AI Works (Early 2026)

In early 2026, the Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the D.C. Circuit's ruling that purely AI-generated works cannot be copyrighted because they lack human authorship. This is now settled law.

• Practical impact: If you generate text with ChatGPT and do not modify it, the raw output has no copyright protection. Anyone can copy and republish it without consequence.
• Human selection and arrangement CAN establish copyright: However, if you select, arrange, and edit AI outputs in a creative way—adding your own commentary, structuring a compilation, or substantially revising the content—your contributions are protectable.
• The "authorship threshold": Courts have not yet defined how much human contribution is required. The Copyright Office's 2025 guidance suggests that prompt engineering alone is not enough.

5. Copyright Office Guidance Update: Human Authorship Required (2025)

The U.S. Copyright Office issued updated guidance on AI-generated works in three installments in 2025. Here's what changed:

January 2025 Report (Part 2): "Human Authorship Is a Bedrock Requirement"

• Reaffirmed that prompt engineering alone does not establish authorship.
• Clarified that AI-assisted works are copyrightable only to the extent they contain "human-authored expression."
• Recommended that Congress consider legislation to clarify AI copyright issues (no action taken as of March 2026).

February 2025: First AI-Assisted Image Registered

The Copyright Office registered its first AI-assisted visual work: a graphic novel illustration created with Midjourney. The applicant documented 35 iterative human edits including color adjustments, compositional changes, and manual digital painting. The Office registered the work but excluded the AI-generated base layer from copyright protection.

• Lesson for ChatGPT users: Document your creative process. Save drafts showing how you modified AI output. This evidence could support a copyright registration application.

May 2025 Report (Part 3): Fair Use Is Case-by-Case

The Copyright Office concluded that whether AI training constitutes fair use depends on the facts of each case. Key factors:

• Commercial vs. research purpose: Training for a commercial product (like ChatGPT) weighs against fair use.
• Amount used: Training on entire copyrighted works weighs against fair use.
• Market harm: Whether the AI output serves as a market substitute for the copyrighted work.
• Transformative purpose: Whether the use creates something new or merely replaces the original.

The Copyright Office did not endorse a blanket rule that AI training is fair use. Instead, it recommended courts apply traditional fair use analysis on a case-by-case basis.

6. Fair Use Scorecard: 3 Judges, Split Decisions

In 2025, three federal judges issued substantive rulings on whether AI training is fair use. The results are mixed—and contradictory:

Case	Court	Date	Ruling	Key Reasoning
Thomson Reuters v. Ross Intelligence	D. Del.	Feb 2025	✗ AGAINST AI	Training on legal database harmed the market for the original work; not sufficiently transformative
Bartz v. Anthropic	N.D. Cal.	Jun 2025	⚠ MIXED	Training on legally-acquired books = fair use; training on pirated copies = NOT fair use
Kadrey v. Meta	N.D. Cal.	Apr 2025	✓ FOR AI	Training on books was "quintessentially transformative"; no market harm because LLM is not a book substitute

7. Settlement Wave: Anthropic Pays $1.5 Billion in Bartz v. Anthropic (June 2025)

Despite the mixed ruling in Bartz v. Anthropic, Anthropic settled the case for $1.5 billion in June 2025—the largest AI copyright settlement to date. The settlement included:

• A licensing agreement allowing Anthropic to continue using the plaintiffs' works in training data
• Payments to a class of affected authors (approximately 5,000 writers)
• No admission of wrongdoing by Anthropic

Other settlement activity in 2025-2026:

• OpenAI settlement negotiations: Multiple reports indicate OpenAI is in settlement talks with plaintiffs in several consolidated cases. No public terms have been disclosed as of March 2026.
• Trend toward licensing: Major AI companies are increasingly signing licensing deals with publishers, record labels, and media companies rather than litigating fair use claims. See AI Output Rights Hub for ongoing updates.

8. Broader AI Copyright Litigation Landscape (2025-2026)

As of early 2026, over 70 copyright lawsuits have been filed against AI companies. Notable developments:

Carreyrou et al. v. Anthropic, OpenAI, Google, Meta, xAI, Perplexity (December 2025)

Investigative journalists including John Carreyrou (author of Bad Blood) sued six major AI companies, alleging training on "pirated books" obtained from illegal repositories like Library Genesis and Sci-Hub. Key allegations:

• AI companies knowingly used pirated training datasets
• Training data included books that were never legally acquired
• Defendants did not attempt to verify the legality of training sources

This case could undermine fair use defenses if plaintiffs prove the training data was obtained through copyright infringement.

UMG/Concord v. Anthropic (January 2026): $3.1 Billion Music Copyright Claim

Universal Music Group and Concord Music filed a $3.1 billion lawsuit against Anthropic in January 2026, alleging that Claude was trained on over 20,000 copyrighted songs without permission. The complaint alleges:

• Claude can generate lyrics that are substantially similar to copyrighted songs
• Anthropic's training data included lyrics scraped from unauthorized sources
• No licensing agreements were in place for musical works

This is the largest music copyright claim ever filed against an AI company.

Disney/Universal v. Midjourney (June 2025): Visual Works Litigation

Disney and Universal Pictures sued Midjourney, alleging that the image generation AI was trained on copyrighted film stills, promotional materials, and character designs. The case is ongoing as of March 2026.

What This Means for ChatGPT Users

Here's my assessment as a California attorney who closely follows AI litigation:

1. Output ownership is yours—but copyright protection is uncertain

   OpenAI's Terms assign you ownership of outputs, and that contractual right is enforceable against OpenAI. But third parties are not bound by OpenAI's Terms. If someone else generates identical ChatGPT output, you cannot stop them without copyright—and purely AI-generated content is not copyrightable.

2. Add substantial human contribution to establish copyright

   Based on the Copyright Office's 2025 guidance and early case law, you should:

   • Substantially edit AI output (not just grammar fixes)
   • Add original analysis, commentary, or creative expression
   • Use AI as a tool in a larger creative process, not as the sole author
   • Document your creative process (save drafts, prompts, and edits)

3. Opt out of training if using for business

   If you're using ChatGPT Plus for commercial work, opt out of data training in Settings > Data controls. If you're using the API or Enterprise tier, training is off by default. The January 2026 discovery order makes clear that OpenAI user conversations are potentially subject to legal discovery—protect confidential information accordingly.

4. Fair use is still unresolved—proceed with caution

   The split decisions in 2025 mean we do not yet have clear legal guidance on whether AI training is fair use. Until appellate courts (or the Supreme Court) resolve this question, AI companies face ongoing litigation risk—and users face uncertainty about whether outputs might inadvertently reproduce copyrighted content. Always verify outputs for potential infringement before commercial use.

5. Expect licensing agreements to become standard

   The settlement wave suggests that AI companies are moving toward a licensing model for training data. This may reduce legal uncertainty in the long term, but it also means that AI services may become more expensive as companies pay for content licenses.