Intellectual property: training data, outputs, patents and trade secrets

Generative AI has put intellectual property law at the center of AI governance. Models are trained on vast collections of text, images, music and code, much of it protected by copyright; they produce outputs that may resemble existing works; and they raise new questions about who, if anyone, owns what they create. Courts in several countries are deciding cases that will shape the industry, and legislatures and copyright offices are studying reforms. Beyond copyright, AI raises questions about patents for AI-assisted inventions, trade secret protection for models and data, and the use of people's voices and likenesses. This lesson covers copyright in training data, copyright in outputs, infringing outputs and how to manage them, patents, trade secrets, personality rights, and licensing of models and data.

Copyright basics

Copyright protects original expression (literary, artistic, musical and dramatic works, including software) but not ideas, facts or methods. It gives the owner exclusive rights to reproduce, adapt, distribute and communicate the work, subject to exceptions. Training a model involves making copies of works (to collect, clean and process them), and a model's outputs may reproduce or adapt them, so both stages can engage copyright. The key legal questions are whether training copies fall within an exception, whether models themselves contain copies, and whether particular outputs infringe.

Training data: exceptions and litigation

In the United States, training on copyrighted works without permission is defended under fair use, which courts assess using four factors: the purpose and character of the use (including whether it is transformative and commercial), the nature of the work, the amount used, and the effect on the potential market for the work. Several important decisions have emerged. In Bartz v. Anthropic (June 2025), a federal judge held that training a large language model on lawfully acquired books was fair use, as a highly transformative use, but that downloading and keeping pirated copies in a central library was not; the case then settled for 1.5 billion dollars. In Kadrey v. Meta (June 2025), another judge found fair use on the record presented but emphasized that market dilution, from AI-generated works competing with the originals, could defeat fair use in other cases. In Thomson Reuters v. Ross, concerning a non-generative legal research tool trained on Westlaw headnotes, the court held that the copying was not fair use, and in September 2026 the Third Circuit affirmed, the first appellate ruling on AI training, finding harm both to the legal research market and to an emerging market for licensing training data, while expressly leaving generative AI questions open. Many other cases, including those brought by news organizations, authors, artists and music publishers, are pending. The US Copyright Office published reports in 2025 on copyrightability of AI outputs and on training, the latter suggesting that some training uses are fair and others, especially those producing competing outputs from pirated or licensed-market content, may not be.

The European Union has text and data mining exceptions under its 2019 Copyright in the Digital Single Market Directive: one for research organizations, and a general one that applies unless rights holders have reserved their rights in a machine-readable way (an opt-out). The EU AI Act requires general-purpose AI model providers to have a policy to respect these opt-outs and to publish a summary of training content (Lesson 3). Japan and Singapore have broad exceptions for computational analysis. The global picture is therefore fragmented, and developers increasingly license data from publishers, image libraries, music companies and platforms, partly to reduce legal risk and partly to access higher-quality data.

Copyright in outputs

Can AI-generated outputs be protected by copyright, and if so, who owns them? In most jurisdictions, copyright requires human authorship.

In the United States, the Copyright Office registers only works of human authorship. In Thaler v. Perlmutter (2025), the D.C. Circuit upheld the refusal to register an image generated autonomously by an AI system. The Office's 2025 report on copyrightability concluded that prompts alone generally do not provide sufficient human control over expressive elements, but that human selection, arrangement and modification of AI-generated material, and human-authored elements in combined works, can be protected; applicants must disclose and disclaim AI-generated content beyond a minimal amount.

Practical consequences follow. Content generated purely by AI may be freely copied by others; organizations that need exclusive rights (for marketing, entertainment or software) should ensure meaningful human creative contribution and document it. Provider terms of service usually assign whatever rights exist in outputs to the user, but cannot create rights the law does not grant.

Infringing outputs

Regardless of how training is treated, an output that reproduces a substantial part of a protected work, or is substantially similar to it, can infringe, and the user who publishes it may be liable as well as, potentially, the provider. Models can memorize and reproduce training text, code, images (such as famous artworks or characters) and lyrics, especially for works duplicated many times in training data. Providers mitigate with deduplication of training data, filters that block reproduction of known protected content, and refusals for requests to reproduce works. Users mitigate by checking outputs, especially code and published content.

# Checking generated text for near-verbatim reproduction of a protected source, using word n-gram overlap
def shingles(text, n=6):
    w = text.lower().split()
    return {tuple(w[i:i + n]) for i in range(len(w) - n + 1)}
source = ("the old lighthouse keeper climbed the spiral stairs each evening to light the lamp "
          "that guided the fishing boats home through the narrow rocky channel before the storms of autumn")
outputs = {
    "close paraphrase": "each evening the old keeper of the lighthouse climbed the winding stairs to light the lamp for the boats",
    "verbatim passage": "she remembered how the old lighthouse keeper climbed the spiral stairs each evening to light the lamp that guided the fishing boats home",
    "original text": "a ferry crossed the bay at dawn while gulls circled above the fish market",
}
src = shingles(source)
for name, text in outputs.items():
    sh = shingles(text)
    overlap = len(sh & src) / max(len(sh), 1)
    flag = "review for copying" if overlap > 0.3 else "low overlap"
    print(f"{name:18s} 6-gram overlap with source {overlap:5.1%}  {flag}")

close paraphrase   6-gram overlap with source  0.0%  low overlap
verbatim passage   6-gram overlap with source 83.3%  review for copying
original text      6-gram overlap with source  0.0%  low overlap

The verbatim passage shares 83 % of its six-word sequences with the source and should be reviewed, while the original text shares none. The close paraphrase also shares none, which shows the limitation of simple n-gram checks: copyright infringement can occur through substantial similarity of expression without exact copying, as with paraphrases of distinctive passages, and detecting it requires human judgment. Similar tools exist for code, where license compliance tools compare generated code against open-source repositories to detect copied code carrying license obligations.

Many AI vendors now offer copyright indemnities to business customers, promising to defend and compensate them for claims that outputs infringe, subject to conditions such as using the vendor's content filters and not deliberately trying to reproduce protected works. Organizations should read these conditions carefully (Lesson 11).

Patents and AI

Patents protect inventions that are new, non-obvious (inventive) and useful, for a limited term in exchange for public disclosure. Two AI issues arise. First, inventorship: courts and patent offices in the United States, the United Kingdom, the EU and elsewhere, in the DABUS cases, have held that an inventor must be a natural person, so an AI system cannot be named as inventor; Canadian patent authorities have taken the same position. Inventions made with AI assistance can be patented if a human made a significant contribution to the conception of the claimed invention, though the guidance on what counts has evolved and should be checked. Second, patentability of AI itself: algorithms and mathematical methods are not patentable as such, but AI-based inventions that produce a technical effect or a practical application, such as improved image processing or control of a physical process, may be patentable, with jurisdictions drawing the line differently. Organizations must also weigh patents against trade secrecy: patents require disclosure, while much AI value lies in data, weights and know-how that are easier to keep secret.

Trade secrets and confidential information

Trade secret law protects commercially valuable information kept secret through reasonable measures. Model weights, training data, data pipelines, prompts and evaluation sets are often protected this way, and their protection depends on security, access controls, confidentiality agreements and careful contracts with partners. Two common risks deserve attention. First, employees pasting confidential information into public AI tools may disclose it to the provider and lose trade secret protection; several companies restricted employee use of public chatbots after incidents of this kind, and many now provide enterprise tools with contractual protections. Second, outputs of a model can sometimes reveal training data or system prompts, so confidential data used in fine-tuning or retrieval needs safeguards (AI-430). Model extraction, where an attacker queries a model to train a copy, raises both technical and legal questions; providers' terms usually prohibit using outputs to develop competing models.

Voices, likenesses and personality rights

Generative AI can imitate a person's voice, face and style, raising questions beyond copyright, which does not protect a voice or a style as such.

In the United States, state right of publicity laws protect against unauthorized commercial use of a person's name, likeness and, in many states, voice. Tennessee's ELVIS Act (2024) explicitly covers AI voice cloning, and other states have followed. The federal TAKE IT DOWN Act (2025) addresses non-consensual intimate deepfakes, and many states regulate deceptive deepfakes in elections. Performers' unions have negotiated contract terms requiring consent and compensation for digital replicas.

Licensing models and data

Much AI is built on licensed components. Open models are released under licenses ranging from permissive open-source licenses (such as Apache 2.0 or MIT) to custom licenses that restrict certain uses, require attribution, or limit use by very large companies; some "open-weight" licenses are not open source in the traditional sense. Datasets carry their own licenses and terms. Before using a model or dataset, organizations should check the license for permitted uses (commercial use, modification, redistribution), attribution and notice requirements, restrictions on use cases, obligations passed to downstream users, and any terms about using outputs to train other models. Keeping a record of components and their licenses, a software or AI bill of materials, supports compliance and due diligence.