Free tools Windows power users keep installed
One-click scans. No signup required.
Generative AI has enabled an extraordinary industrial-scale use of creative and informational works, often without creators negotiating permission or compensation first. But calling it “the most brazen intellectual-property theft in history” is a forceful moral judgment, not a settled legal or historical finding. The distinction matters: copying works to train a model, obtaining them from unauthorized sources, retaining a searchable archive, and generating outputs that reproduce protected expression can raise different legal questions.
A landmark U.S. book case shows why no single slogan captures the issue. On July 20, 2026, a federal court approved Anthropic’s $1.5 billion settlement over claims involving books obtained from pirate repositories. That resolved claims within that case; it did not establish that all AI training infringes copyright. The court’s final approval order must be read alongside the still fact-specific law governing training and outputs.
What “intellectual-property theft” means in an AI dispute
“Theft” communicates a real grievance: creators may see companies turning their work into valuable commercial systems without asking or paying them. Legally, however, the word is imprecise. Copyright infringement generally concerns unauthorized acts such as copying, distributing, adapting, performing, or displaying protected expression; it is not simply the taking of a physical object.
Copyright is only one possible issue. AI systems can also raise questions involving trademarks, rights of publicity, trade secrets, patents, contracts, and moral rights. The relevant claim depends on the work, the conduct, the jurisdiction, and the way a model or its output is used. A generated logo that falsely implies brand endorsement, for example, presents a different issue from a model trained on a copyrighted book.
#1 Best Overall
This article’s legal discussion is primarily about U.S. copyright law. The United States’ fair-use analysis is not a worldwide rule: the EU, UK, Canada, Japan, and other jurisdictions have different text-and-data-mining exceptions, transparency rules, licensing systems, and moral-rights protections. A use that has one legal analysis in one country may face another when a system is deployed or an output is distributed elsewhere.
Why AI makes the copying dispute different in scale
Copying is not new. The distinctive concern is the industrial conversion of vast numbers of individually created works into general-purpose commercial infrastructure. A typical development pipeline may gather material from the web or other sources, make working copies, normalize text or images, train or fine-tune a model, and then deploy that model in products. The particulars vary, and complete training-set disclosure is not universal.
- Volume and speed: Automated collection and processing can operate at a scale difficult to replicate through individual human reading or licensing negotiations.
- Reuse: A dataset or trained capability can support multiple products, customers, and later systems.
- Opacity: When sources and data-handling practices are not fully disclosed, a creator may struggle to learn whether a work was included or how it was used.
- Economic leverage: Aggregated works can help build products that compete in markets for writing, images, music, code, journalism, and reference material.
- Detection: Once training has occurred, identifying every source or removing every possible influence may be difficult. That does not mean the model necessarily contains a conventional copy of each source file.
The U.S. Copyright Office has identified licensing, attribution, compensation, and creator incentives as central policy questions, while noting practical difficulty in tracking or paying very large numbers of rights holders. Its economic implications report describes those challenges; scale alone, however, does not decide whether a particular use is lawful.
Three layers to examine: input, model, and output
“Was the AI trained on my work?” is only the beginning of a useful inquiry. A dispute can involve distinct conduct at each layer:
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match- Input: Was the work protected, and how was it obtained? Was it licensed, public domain, supplied by a user, scraped from an accessible site, or downloaded from an unauthorized repository? Public accessibility is not the same as permission or public-domain status.
- Model: What copies were made during processing? Were they retained, searchable, or used for training? Could the model reproduce a substantial expressive portion of a source?
- Output: Does a response reproduce protected expression, a recognizable character or logo, or a substantial part of a work? Does it compete with or substitute for the original?
Permission, license terms, opt-outs, contracts, commercial use, and the surrounding facts can matter at more than one layer. An argument that training is transformative does not automatically answer whether a source was lawfully acquired or whether an output infringes.
What the major disputes establish—and what they do not
| Dispute | Works and conduct at issue | What the available record establishes | What it does not establish |
|---|---|---|---|
| Bartz v. Anthropic | Books, including claims involving downloads from LibGen and PiLiMi and use of books in AI development. | A federal court’s July 20, 2026 final order approved a $1.5 billion settlement plus interest. The Works List contained 482,460 works. The order reported that 91.3% had been claimed as of April 16, 2026, and estimated roughly $3,000 per work, subject to deductions, allocation, and claim validity. See the final approval order and the preliminary approval order. | The settlement is not a ruling that all AI training is infringement, nor does the estimated per-work amount mean every rights holder received that amount. It resolves claims within the settlement’s scope. |
| OpenAI consolidated copyright litigation | Claims involving books, journalism, and other works, with disputes that include training-related data and records. | A 2026 discovery order addressed production and discussion involving large data reservoirs and training-related records. See the discovery order. | A discovery ruling is not a merits judgment. The cited order does not establish liability across the pending claims. |
| Image-model disputes | Artists and image rights holders have raised allegations involving image datasets, generated images, and recognizable entertainment properties, including disputes involving Stability AI, Getty Images, and Midjourney. | The central questions can include source copying, substantial similarity, and whether a particular output reproduces a protected image, character, logo, or composition. | This article does not establish a general ruling that image-model training is lawful or unlawful, or that visual style by itself is protected by copyright. |
| Software and code | Questions include training on public repositories, license conditions, and outputs that reproduce code. | Publicly available code can still carry copyright and license terms. Verbatim or near-verbatim output can raise different questions from learning general programming patterns. | Public access alone does not establish permission for every commercial use, and the facts of an output and applicable license matter. |
The U.S. Copyright Office’s Part 3 report on generative-AI training and the Congressional Research Service overview describe a developing, fact-specific field. Neither supports a categorical statement that every training use is lawful or every one is unlawful.
Why the Anthropic settlement is important, but not a verdict on all training
The Anthropic book dispute makes the input-versus-training distinction concrete. A court’s treatment of training on lawfully acquired books was not a blanket license to obtain books from any source or keep any copy for any purpose. The case separately involved allegations and findings concerning books downloaded from pirate repositories and retained. Those acquisition and retention issues are not interchangeable with the question of whether training itself can qualify as fair use.
The settlement’s scale is consequential: the approved fund was $1.5 billion plus interest, and the order identified 482,460 works on the Works List. Its estimated amount of about $3,000 per work was not a guaranteed individual payment; valid claims, deductions, allocation, and administration affect distributions. Most importantly, a settlement resolves claims without producing a universal judicial rule for other companies, datasets, or training methods.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteHow U.S. fair use applies to AI training
U.S. fair use is assessed under four statutory factors, considered together rather than as a checklist where one favorable factor decides the case:
- Purpose and character: Courts consider the use’s purpose and whether it is transformative. Commercial purpose can matter, as can whether a new use serves a different function or instead replaces the source’s role.
- Nature of the work: Factual material and highly creative expression can weigh differently. Many books, images, songs, and other works used in training are creative.
- Amount and substantiality: Both the quantity copied and the qualitative importance of what was taken may matter. A system may use an entire work, but that fact does not mechanically resolve the factor.
- Market effect: Courts consider harm to the original and relevant potential markets. Outputs that substitute for a work, or for a licensing market, may be significant to this analysis.
AI complicates all four. Training may have a new technical purpose, yet use whole creative works; the model may not ordinarily return those works, yet particular prompts may elicit expressive passages; and a tool may assist users while also competing with creators. Courts may need to distinguish copies made to assemble a dataset from model training and from particular outputs. The Copyright Office and CRS both describe the results as dependent on circumstances, not a settled industry-wide answer.
Memorization is a real risk, but not proof that every model is a copy
A model need not preserve a normal PDF or image file in order to raise concerns about reproduction. Technical researchers use “memorization” to describe cases where a model can reconstruct a near-exact portion of a training example. The study “The Files are in the Computer” examines memorization and extraction in language models.
Possible symptoms include repeated book passages, long lyrics, distinctive image details, or code appearing verbatim. Researchers and users may probe models with specific prompts, including extraction attempts. But resemblance is not automatically memorization: a generic idea, broad style, or common phrase may resemble a source without reproducing protected expression. Conversely, showing that a model can emit a particular passage does not, by itself, establish that the entire model is an unlawful copy. The source, amount, context, acquisition, and output use still matter.
Different creative fields raise different risks
Books, journalism, and reference material
Authors and news publishers have challenged uses of written works in training and in systems that answer questions or summarize content. An AI answer may affect rights holders after the training stage if it reproduces passages, captures the value of a reference product, or gives users a substitute for visiting an article. Whether a specific answer is infringing depends on what it reproduces and how it functions in the market. Search and plagiarism-detection cases have sometimes treated mass copying as transformative, but that does not automatically resolve generative systems that produce substitutive outputs; see the Copyright Office’s training report.
Images, characters, brands, and a person’s identity
A prompt that asks for a broad visual mood is not the same as an output that reproduces a particular protected illustration, composition, character, or logo. Copyright does not generally grant ownership of an abstract style as such, but a specific image can still infringe. Separate questions may arise under trademark law, trade dress, or rights of publicity if generated material uses a brand or imitates an identifiable person’s likeness or voice in a misleading or unauthorized way.
Code and software licenses
Code may be protected even when posted publicly, and repository licenses can impose conditions such as attribution or share-alike obligations. A model’s general ability to write a programming pattern is not the same as returning a substantial block of a particular project’s code. For businesses, verbatim output can also introduce security, provenance, and license-compliance problems independent of whether the model was trained on public repositories.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The strongest case for AI developers—and the strongest case for creators
Why developers argue training can be lawful
- Training extracts statistical patterns or supports a new function rather than distributing source copies to users.
- Models do not necessarily store source works as ordinary retrievable files.
- Licensing every item in huge corpora may be difficult or costly, potentially limiting research and useful products.
- AI can support accessibility, productivity, search, and new forms of expression.
- Outputs are not necessarily substantially similar to any one training work.
These arguments can carry legal weight in particular cases; they are not a substitute for examining the source, copying, retention, model behavior, outputs, and market effects.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Why creators object
- The initial act of making copies may matter even if the deployed model is not a file archive.
- Public availability is not the same as a license, and some disputes involve alleged use of unauthorized sources.
- Outputs may reproduce protected expression or compete with the market for the source.
- Creators may have little visibility into datasets or practical ability to negotiate before use.
- Commercial systems can capture aggregate value from many works while leaving individual creators with uncertain attribution, compensation, or control.
The Copyright Office’s economic analysis discusses the difficulty of attribution and compensation at scale, as well as possible effects on incentives to create. Neither innovation benefits nor creator harms answer every legal claim by themselves.
Licensing and provenance offer alternatives to a permission-or-progress stalemate
AI systems can be developed with public-domain material, licensed works, permissioned datasets, or combinations of sources. Licensing may be direct, collective, or mediated through libraries and marketplaces. More transparent provenance can help buyers and rights holders understand what a dataset contains and which uses were authorized.
Potential arrangements include opt-in permissions, opt-out signals, compensation pools, attribution systems, contractual warranties, and mechanisms to filter or remove material. Each has trade-offs: opt-out places the burden on creators to find and flag uses; licensing can be costly to administer; and a provenance record does not itself prove that every copy or output is lawful. The Copyright Office identifies licensing, compensation, and tracking as significant policy questions, not as problems already solved by a single mechanism.
What creators and businesses can do now
For creators and rights holders
- Keep dated source files, publication records, contracts, and license terms that help establish ownership and provenance.
- Review the terms of platforms where you publish, including any permissions granted for hosting, indexing, or machine-learning uses.
- Use available opt-out or licensing processes where they fit your goals, while recognizing that effectiveness depends on the systems and datasets involved.
- Document suspected outputs with the prompt, date, model or service, and complete result; similarity tools can help locate material but do not conclusively prove infringement.
- Before sending a legal notice or alleging infringement publicly, assess the specific work, output, ownership, and applicable rights with qualified counsel.
For companies procuring or deploying AI
- Ask vendors what is known about training-data provenance, retention, and applicable licenses; record the answers and the service/version reviewed.
- Review contractual warranties, indemnity scope and exclusions, audit rights, data-retention terms, and claim procedures rather than relying on a general “commercial use” label.
- Set rules against prompts intended to reproduce named works, characters, brands, or proprietary code.
- Use output review and filtering for commercial publishing, and keep a process for documenting sources and responding to rights complaints.
- Separate the question “may we use this tool?” from “may we publish this particular output?” Contractual permission from a vendor does not automatically clear every third-party right.
So, is it the most brazen IP theft in history?
The strongest supportable answer is narrower and more useful than a simple yes or no: generative AI has made possible an unusually vast, rapid, and economically consequential mass ingestion of creative and informational works. The industrial conversion of countless works into commercial infrastructure is a serious challenge to creators’ bargaining power and to existing copyright frameworks.
But “the most brazen intellectual-property theft in history” remains a rhetorical characterization. Scale does not settle fair use; a major settlement does not prove universal infringement; and unlawful acquisition, training, retention, memorization, and output substitution are distinct questions. Courts and policymakers are still defining where the lines fall, and the answer will depend on the specific works, data sources, model behavior, outputs, markets, and jurisdiction.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
