The legal risks of training AI on scraped data.

What does the law say about scraping data for AI training?

In the US there is no statute on AI training, so courts apply fair use one case at a time, and they have split on the facts. On 23 June 2025 Judge Alsup held in Bartz v. Anthropic that training on books was "exceedingly transformative" and fair use, and that scanning books Anthropic had bought was fair use too. He did not extend that to more than seven million books downloaded from pirate libraries, which he sent toward trial. Two days later, in Kadrey v. Meta, Judge Chhabria ruled for Meta but said the result reflected the authors' thin evidence of market harm, not a finding that such training is lawful in general.

The first appellate ruling went the other way. On 29 September 2026 the Third Circuit held that Ross Intelligence's use of 2,243 Westlaw headnotes to train a competing legal research tool was not fair use, while noting that Ross's tool could not generate new text. The US Copyright Office's May 2025 pre-publication report on AI training, which is not binding, said that using copyrighted works to produce competing content, "especially where this is accomplished through illegal access, goes beyond established fair use boundaries."

The EU works differently. Article 4 of the 2019 Copyright in the Digital Single Market Directive allows text and data mining of lawfully accessible content unless the rightsholder has reserved its rights, for online content in a machine-readable way. This article is a summary of public sources, not legal advice; have counsel review any dataset before you train on it.

What does the EU AI Act require for training data?

Since 2 August 2025, Article 53 of the AI Act requires every provider of a general-purpose AI model placed on the EU market, open-source ones included, to do two things about training data:

• Keep a policy to comply with EU copyright law, including identifying and honouring opt-outs made under Article 4(3) of the Directive
• Publish a summary of the training content using the European Commission's template, which for scraped data asks for the crawlers used, the collection period and the top 10% of domains scraped
• Models already on the market before 2 August 2025 have until 2 August 2027 to publish their summary
• The AI Office can enforce from 2 August 2026, with fines up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher

In practice, a provider has to be able to say where its training data came from and show that opt-outs were respected. That is hard to do after the fact for a large crawl.

Which AI training data lawsuits matter?

Status as of 20 October 2026. Several of these are still pending; see sources for each.

What does it cost to get training data wrong?

• Settlements: Bartz v. Anthropic settled for $1.5 billion, about $3,000 per work, and Anthropic must destroy the files it took from the pirate libraries
• Model deletion: the FTC ordered Everalbum (2021) to delete face recognition models built from users' photos, WW International's Kurbo (2022) to destroy algorithms built from children's data and pay $1.5 million, and Rite Aid (2023) to delete images and the algorithms developed from them
• Injunctions: Reddit's suit against Perplexity and three scraping firms seeks a court order to stop the scraping as well as damages
• Rework: if a source has to be removed, every model trained on it may need to be retrained, which costs compute and time

Can website terms of service make scraping a breach of contract?

Yes, and contract claims do not depend on copyright. In LinkedIn v. hiQ, the court held in November 2022 that LinkedIn's user agreement ban on scraping was enforceable, and hiQ agreed to a $500,000 judgment and an injunction against further scraping. Reddit's June 2025 suit against Anthropic is built on the same kind of claim: breach of its user agreement, trespass on its servers and interference with its privacy commitments to users. A federal court sent it back to California state court in March 2026. Reddit's October 2025 suit also names the scraping and proxy firms that sold data on, so buying scraped data from a vendor does not take the buyer out of the dispute.

What privacy risks come with scraped data?

Scraped data almost always contains personal data, and in the EU the GDPR requires a lawful basis for processing it whether or not it was public. The Dutch Data Protection Authority fined Clearview AI EUR 30.5 million in 2024 for building a database of more than 30 billion photos scraped from the internet without a lawful basis, and warned that Dutch organisations using the service could be fined too. Italy's Garante fined OpenAI EUR 15 million in December 2024 over the legal basis for training ChatGPT on personal data; the Court of Rome upheld OpenAI's appeal in a judgment published on 18 March 2026. The outcomes vary, but the investigations alone take years and the questions are the same: where the personal data came from and on what basis it was used.

Is scraped data a quality problem as well as a legal one?

• Contamination: benchmark questions and answers that leak into crawled data inflate evaluation scores, as surveyed by Xu et al. (2024)
• Synthetic content: a 2024 Nature paper found that training on recursively generated model output causes models to lose the rare parts of the original data
• No edge: every lab can crawl the same public web, so it does not set one model apart from another
• Gaps: much of what a model needs to learn, such as first-person video of real tasks, is not on the open web at all

What changes with licensed data and per-record provenance?

With licensed data, the right to train comes from a contract with the party that holds the data, rather than from a fair use argument made after the model ships. When each record carries its source and license, a lab can fill in an EU training summary from its own records, answer a rightsholder or regulator, and remove one source without guessing what else it touched.

That is how Indri Labs works. We scope a request for free and send a brief with every source named and priced, a sample, and the total cost with our fee on its own line. On approval you sign one contract, with us. We license the data from each company that holds it rather than scraping it, and we record new first-person video where none exists. We remove duplicates and personal data, check everything against your spec, and deliver one dataset with the source and license for every record, ready for your legal review. More on how Indri Labs works.

What should a lab ask before buying training data?

• Who holds the rights to each source, and can you see the license?
• Was any of it scraped, and if so, against whose terms of service?
• Were text and data mining opt-outs checked for EU content?
• Has personal data been removed, and on what basis was any of it collected?
• Does every record carry its source and license, so one source can be removed later?
• Has the data been checked for duplicates and for overlap with public benchmarks?
• Is there enough source detail to complete an EU AI Act training summary?