Skip to main content

Who owns the output: IP, training-data provenance, and the liability chain when AI work goes wrong

The short answer

"Who owns the output" is three questions, not one

Did the input infringe, is the output ownable, does the output infringe a third party. They have different laws, different answers, and different exposed parties. The first expert move is to say which one you are actually facing.

What you will be able to do

  • Separate the three distinct ownership questions that hide inside "who owns the output": does training on the input infringe, is the output itself protectable, and does the output infringe someone else's rights.
  • Trace the training-data provenance of a specific AI system from original source to final output, using the provenance file your organization built earlier (see Topic 2.6).
  • Analyze the liability chain of a deployed AI system, naming each party (foundation-model provider, integrator, deployer, end user) and the basis on which each could be held responsible.
  • Distinguish what a contract allocates (indemnity, warranty, ownership assignment) from what the law imposes (copyright, product liability, professional duty) and from what the evidence can actually prove.
  • Evaluate a vendor's intellectual-property indemnity for the carve-outs that decide whether it protects you when a claim actually lands.
  • Apply the reasoning of Thomson Reuters v. Ross Intelligence and the 2025 United States Copyright Office guidance on human authorship to your own organization's outputs without overstating what either one settled.
  • Produce a plain-language ownership-and-liability map for one AI system that a successor, an auditor, or opposing counsel could read and follow.

The ideas, one by one

Training-data provenance is now legal exposure, not hygiene

Thomson Reuters v. Ross Intelligence (D. Del., February 2025) showed that training on someone else's protected content, to build a competitor, without a license, can be infringement and not fair use. The EU AI Act (Article 53) now requires GPAI providers to publish a training-content summary. The provenance file you built in Module 2 is your only way to answer the input question (see Topic 2.6).

Ross decided one narrow thing; do not overstate it

It was non-generative, commercial, and directly competing. It did not rule on generative AI, and it is on appeal. Telling a board that "AI training is illegal" is a falsification. State exactly what was held and what remains open.

Prompts do not earn you copyright

Under the 2025 United States Copyright Office guidance, raw prompt output is not protectable; only recorded human creative contribution is. If you want to own your AI-assisted work, engineer and record a real human editorial step. That same step also lowers your infringement risk.

Output ownership changes by jurisdiction

The United States denies copyright to prompt-only output; the Beijing Internet Court granted it to an AI-assisted image (see Topic 10.6). A multinational maps ownership per country, not once.

Liability attaches along a chain, on more than one basis

Foundation-model provider, integrator, deployer, end user, each reachable through copyright, product liability, professional duty, or misrepresentation. The deployer who ships the output is usually the most exposed. A tool never absorbs a professional's duty of care.

Contract, law, and evidence are three layers that can disagree

Law sets the baseline; contracts allocate risk between the parties who signed; evidence decides what you can actually prove. A right or a defense you cannot document is worth little, which is why this lives in the evidence module.

Read every indemnity for its carve-outs

Vendor IP indemnities (Microsoft, OpenAI, Adobe, Google, AWS) are real but conditional: safety features on, paid tier, output claims only, capped payout. Record the conditions and whether you can prove you meet them.

Strict liability raises the value of your records

The revised EU Product Liability Directive brings software and AI under no-fault liability with claimant-friendly presumptions. When fault is no longer the question, your logs, warnings, and design records become the decisive evidence (see Topic 10.2).

Exposure outlives the system

Ross shut down in 2021; the ruling landed in 2025. Retiring a system does not retire its liability, so preserve the evidence rather than deleting it. Your finished map feeds the evidence annex and the capstone dossier (see Topic 10.6) (see Topic 13.1).

Name which kind of output infringement you fear

Regurgitation of training data is defended upstream (vendor assurances, anti-memorization safeguards); independent similarity to a protected work is defended downstream (human review, records of independent creation). Aim your controls at the right end of the chain.

The input side has an ownership question too

What your organization puts into a model can cost you IP: confidential material entered into an uncontrolled tool can lose its trade-secret protection. Govern inputs by tool and tier, and record that the policy was followed.

"We only deploy" is not a permanent shield

Under EU AI Act Article 25, renaming, substantially modifying, or repurposing a system can turn a deployer into a provider with heavier duties. Record the role you actually occupy for each system.

Code and real people are special cases

AI-generated code can carry open-source license obligations (scan it before shipping); AI output about a real person can trigger defamation, right-of-publicity, and data-protection claims that have nothing to do with copyright (never ship it without human sign-off).

Listen to it

The same lesson, talked through by two hosts as episode 80 of the podcast.

Real cases

These examples show the three questions and the liability chain in real, documented cases. Each is used here to illustrate one link; the deep case treatment lives with its owning topic where noted. As you read, tag each example to the part of the chain it lights up:

  • The input question: Examples 1, 2, 10.
  • The output-ownership question: Examples 3, 4, 8.
  • The liability chain and its layers: Examples 5, 6, 7, 9.

Reading them this way is itself the skill: given any AI story in the news, an expert first asks which link it is really about.

Example 1: Thomson Reuters v. Ross Intelligence (the input question, decided). The anchor of this topic. A federal court held that training a competing legal-research tool on Westlaw's copyrighted headnotes, sourced through third-party memos after a license was refused, was infringement and not fair use, on the narrow ground of non-generative use for a directly competing commercial product (Thomson Reuters v. Ross Intelligence, D. Del. Feb. 11, 2025). The chain lesson: ownership of the editorial headnotes stayed with Thomson Reuters through every hop (Westlaw editor to LegalEase memo to Ross model). Liability reached the party that commissioned and used the copying, even after that party had dissolved.

Example 2: The New York Times v. OpenAI and Microsoft (the input question, unresolved). The Times sued over training on its journalism without permission, alleging the models can reproduce its articles closely. The case is ongoing and settles nothing yet, but it is the clearest live illustration that a foundation-model provider's data choices create exposure that can reach downstream deployers of that model. This case is the anchor for Module 5.5; here it stands only as the open counterpart to the decided Ross case (see Topic 5.5).

Example 3: The United States Copyright Office human-authorship guidance (the output-ownership question). In January 2025 the Copyright Office confirmed that purely prompt-generated output is not copyrightable, while human creative selection, arrangement, or modification of AI output can be (US Copyright Office, Part 2, 2025). Real consequence for organizations: a media company that wants to own its AI-assisted work must build and keep a record of the human creative contribution, or accept that the raw output sits in the public domain and can be freely copied.

Example 4: The Beijing Internet Court AI-image ruling (output ownership, a divergent jurisdiction). China's first ruling granting copyright in an AI-generated image, based on the prompter's documented intellectual inputs (2023), shows that the output-ownership answer flips across borders. A multinational cannot assume the United States "no copyright in prompts" rule applies in every market. Deep treatment belongs to Module 10.6 (see Topic 10.6); cited here only to make the divergence concrete.

Example 5: Vendor IP indemnities in practice (the contract layer). Microsoft's Copilot Copyright Commitment, OpenAI's Copyright Shield, and Adobe's Firefly indemnity all promise to defend paying customers against third-party copyright claims over covered AI output, and all condition that promise on the customer using the product as intended, with safety features enabled, on paid tiers (Lexology, 2024; Proskauer, 2024). Real consequence: a firm that disabled content filters to get "less restricted" output may have voided the very indemnity it was counting on.

Example 6: The revised EU Product Liability Directive (the liability-basis layer). Directive (EU) 2024/2853, in force from late 2024 and to be transposed by 9 December 2026, brings stand-alone software and AI within strict product liability and eases a claimant's burden of proving defect and causation for complex systems (Hogan Lovells, 2024; Orrick, 2025). Real consequence: an EU deployer of a defective AI feature can face a no-fault claim, which raises the value of every design record, warning, and log to the level of core legal evidence (see Topic 10.2).

Example 7: The AI-insurance retreat (what happens when the chain has no backstop). In 2025, several insurers filed to exclude AI-chatbot and agent liabilities from coverage, signaling that the residual risk at the end of the liability chain may not be insurable on ordinary terms. This is the anchor for Module 8.4 and is referenced here only to close the loop: when contract and insurance both thin out, the deployer's own evidence and controls are the last line (see Topic 8.4).

Example 8: A denied and a partly-cancelled registration (the human-authorship line, drawn in real files). The United States Copyright Office refused to register an AI-generated image that had won a state-fair art prize, on the ground that its expressive elements were produced by a machine rather than a human. The artist has challenged that refusal in federal court, where the case remained pending as of 2026, so the settled point is the Copyright Office's refusal itself, not a final court judgment. In a separate matter, the Office reconsidered the registration for Kristina Kashtanova's comic "Zarya of the Dawn," which combined human-written text and arrangement with AI-generated images, and narrowed the protection to the human-authored elements, excluding the individual AI-generated images. Real consequence for organizations: these are the human-authorship rule from Section 3D applied to concrete works, and they show the line is not theoretical. Recorded human contribution is registrable; the raw machine output is not (US Copyright Office decisions, 2023 to 2025).

Example 9: The EU AI Act value-chain roles (who becomes responsible). Regulation (EU) 2024/1689 places the heaviest obligations on the provider of an AI system and lighter ones on the deployer, but Article 25 provides that a deployer who puts its own name on a high-risk system, substantially modifies it, or repurposes it into a high-risk use takes on the provider's obligations. Real consequence: an organization that white-labels or heavily fine-tunes a vendor's system can move up the liability chain without noticing, which is why a map records the role you actually occupy, not the one on the purchase order (European Commission; AI Act Article 25).

Example 10: The training-content summary as a usable document (provenance, in your hands). From August 2025, providers of general-purpose AI models placing them on the EU market must publish a summary of training content using the AI Office's mandatory template, which asks for model and provider metadata, the main data-source categories (public datasets, licensed datasets, crawled or scraped content, user data, synthetic data), and the copyright and data-protection measures taken (European Commission AI Office template, 24 July 2025). Real consequence: when you deploy a major model, this summary is a real artifact you can request, read, and file against the input question, turning "we do not know what it trained on" into a documented answer or a documented gap.

Where people go wrong

  • "Who owns the output is one question." It is three: did the input infringe, is the output ownable, does the output infringe someone else. They have different laws, different answers, and different exposed parties. Mashing them together is the root error; separating them is the first expert move.
  • "The AI made it, so no human is responsible." The opposite is often true. When AI output causes harm, the law reaches a human or an organization: the deployer who shipped it, the professional who relied on it without the check their duty required, or the provider who supplied a defective product. "The model did it" is not a liability shield; a tool cannot hold a professional's duty of care.
  • "Ross v. Ross Intelligence decided that AI training is illegal." It did not. It held that one company's copying of copyrighted editorial headnotes, to build a directly competing tool, without a license, was infringement and not fair use, and it expressly limited itself to non-generative AI. The generative-AI questions are still open and on appeal. Overstating Ross to a board is a falsification you will be caught on.
  • "We prompted it in detail, so we own the copyright." In the United States, per the Copyright Office's 2025 guidance, prompts alone (however detailed) do not confer copyright in the output. Ownership attaches to human creative selection, arrangement, or modification, and only to those human contributions. Detailed prompting is instruction, not authorship.
  • "Copyright works the same everywhere, so we only need one answer." No. The United States denies copyright to purely prompt-generated output; the Beijing Internet Court granted copyright to an AI-assisted image on the strength of the human's inputs (see Topic 10.6). A multinational must map output ownership per jurisdiction, not once.
  • "The vendor indemnifies us, so we are covered." Only if you meet the carve-outs. Common conditions: you kept the built-in safety filters on, you used a paid and generally available tier, the claim is about the output and not your inputs, and the payout is not capped below the exposure. An indemnity you cannot prove you qualify for is not protection; it is a sentence.
  • "Provenance is an ethics nicety." As of 2025 it is a legal obligation and a litigation defense. EU GPAI providers must publish a training-content summary under Article 53 of the AI Act, and your ability to answer the input question at all depends on the provenance file you built in Module 2 (see Topic 2.6). No provenance, no defense.
  • "Strict liability, fault liability, same thing." Under fault-based rules a claimant must show you were careless; under strict liability (which the revised EU Product Liability Directive extends to software and AI) they need only show a defective product caused harm. The shift to strict liability, with presumptions that ease the claimant's proof, makes your own records the decisive evidence, because your carefulness is no longer the question.
  • "Liability ends when the company or the project ends." Ross Intelligence shut down in 2021; the infringement ruling landed in 2025. Winding down a system does not wind down its exposure, which is one more reason the evidence must be preserved, not deleted, when a system is retired.
  • "The model does not store the training data, so the output cannot infringe." How the model represents information internally is not the legal test. Courts look at what actually came out. A model can produce output substantially similar to a protected work whether or not that work is "stored" anywhere, and regurgitation of near-verbatim training data is a documented phenomenon. Do not offer "it only learns patterns" as a blanket defense.
  • "We only deploy other people's models, so provider duties are not ours." Not necessarily. Under Article 25 of the EU AI Act, a deployer that puts its own name on a high-risk system, substantially modifies it, or repurposes it into a high-risk use inherits the provider's obligations. White-labeling and heavy fine-tuning can quietly move you up the chain. Record which role you actually occupy.
  • "Confidential data is safe to paste into any AI tool because the vendor promises privacy." Read which tier you are on. Enterprise contracts often promise inputs are not used for training; consumer and free tiers frequently make no such promise. And disclosing a trade secret into a system you do not control can destroy its legal secret status entirely, not merely risk a leak. Input governance is an ownership question too.
  • "Regurgitation risk and look-alike risk are the same problem." They are not, and they need different controls. Regurgitation (the model emits near-verbatim training data) is defended upstream, through vendor assurances and anti-memorization safeguards. Independent similarity (the output merely resembles a protected work) is defended downstream, through human review and records of independent creation. A map that does not name which risk it faces aims its controls at the wrong end.

Questions people ask

What is fair use?
A defense in United States copyright law that permits some unlicensed use of a protected work, weighed across four factors (purpose and character of the use, nature of the work, amount used, and effect on the market). In Thomson Reuters v. Ross Intelligence the defense failed because the use was commercial, not transformative, and harmed the market. More on Fair use
What is headnote?
A short editorial summary of a point of law in a court opinion, written by a publisher's editors. The underlying opinions are government works and not copyrightable, but the editorial headnotes can be, which is what Thomson Reuters owned and Ross was found to have copied.
What is transformative use?
In fair-use analysis, a use that adds a new purpose or different character to the original rather than substituting for it. A use that builds a directly competing product, as in Ross, is generally not transformative.
What is training-data provenance?
The documented record of where an AI system's training and input data came from and on what legal basis it could be used. It is the artifact that lets you answer whether training on the input infringed (see Topic 2.6).
What is general-purpose AI (GPAI) model?
Under the EU AI Act, a model trained on broad data that can perform a wide range of tasks and be integrated into many systems. GPAI providers carry specific duties, including publishing a training-content summary. More on General-purpose AI (GPAI) model

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →