Last reviewed:
What is AI training data? Definition and business implications
Training data is the set of texts, images, code, and other content used to train an AI model. Its composition determines what the model knows, what it ignores, its biases, and its legal risks. A major part of current AI litigation concerns their provenance and lawfulness.
Foundation models are trained on corpora of several trillion tokens, whose composition is rarely fully public. Three sources dominate. The public web, via Common Crawl: more than 300 billion archived pages, which constitute the raw material of most LLMs. Book and press corpora, whose lawfulness of use is now contested: the New York Times' complaint against OpenAI (December 2023, still ongoing in 2026) concerns precisely this point. Specifically generated data: RLHF annotations, fine-tuning examples, synthetic data. The traceability of training data has become central. Since August 2025, the European AI Act has required providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training (article 53). On 24 July 2025 the European Commission published a template for this summary, which providers must follow.
Concrete example
The state of litigation in 2026 illustrates legal uncertainty. In June 2025, two US decisions (Bartz v Anthropic, Kadrey v Meta) held that training a model on books could qualify as fair use, being highly transformative. But in Bartz, the judge denied fair use for building a permanent library from pirated books. The case ended in a $1.5 billion settlement (about $3,000 per work), given final approval on 20 July 2026. On the OpenAI side, the flagship New York Times complaint is still ongoing. In November 2025, a judge ordered OpenAI to produce a sample of 20 million de-identified ChatGPT conversations to assess the reproduction of protected content; the order was affirmed on 5 January 2026. The matter is not settled. For a European executive, this US uncertainty adds to GDPR and the AI Act, whose timeline has just been adjusted by the 2026 Digital Omnibus.
FAQ
What is training data and why does it matter?
It is the set of texts, images, code, and other content used to train an AI model. Its composition determines what the model knows, what it ignores, its biases, and its legal risks. A major part of current AI litigation concerns their provenance and lawfulness.
Where does this data come from?
Three sources dominate. The public web, via Common Crawl (more than 300 billion archived pages). Book and press corpora, whose lawfulness of use is now contested. Specifically generated data: RLHF annotations, fine-tuning examples, synthetic data. The composition of these corpora, which run to several trillion tokens, is rarely fully public.
What must a provider publish about its training data?
Since August 2025, the European AI Act has required providers of general-purpose models to publish a sufficiently detailed summary of the content used for training (article 53). On 24 July 2025 the European Commission published a template for this summary, which providers must follow.
Where does copyright litigation stand?
Uncertain. In June 2025, two US decisions (Bartz v Anthropic, Kadrey v Meta) held that training on books could qualify as fair use, but in Bartz the judge denied fair use for a library built from pirated books; the case ended in a 1.5 billion dollar settlement, approved on 20 July 2026. The New York Times' complaint against OpenAI is still ongoing.
Which contractual clauses should you require?
Five. A declaration of the categories of training data used. A guarantee that your business data will not be used to train future versions. The possibility of requesting the logical purge of your data from fine-tuning corpora. A warranty clause in case of a third-party copyright lawsuit. An obligation to inform within 30 days of any substantial change to the corpus.
See also
Further reading
Sources
- The New York Times Company v. Microsoft Corporation, S.D.N.Y. No. 23-CV-11195, filed December 2023, in discovery phase in 2026. https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act), articles 53-55 on foundation model providers' obligations. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement, The Authors Guild, July 2026. https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/
- Fair Use and AI Training: Two Recent Decisions Highlight the Complexity of This Issue (Bartz v. Anthropic, Kadrey v. Meta), Skadden, July 2025. https://www.skadden.com/insights/publications/2025/07/fair-use-and-ai-training
- OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs, National Law Review, January 2026. https://natlawreview.com/article/openai-loses-privacy-gambit-20-million-chatgpt-logs-likely-headed-copyright
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX%3A32026R1744
- Template for the public summary of training content, European Commission (AI Office), 24 July 2025, WilmerHale analysis. https://wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/european-commission-releases-mandatory-template-for-public-disclosure-of-ai-training-data
- Common Crawl, archive overview. https://commoncrawl.org/