Generative AI is changing everything — but at what cost?
Most major models (GPT, Gemini, Claude, LLaMA) were trained on copyrighted books, code, music and videos without clear consent. That’s not a detail: it’s the structural flaw behind today’s “AI boom.”
Billions in potential legal liabilities are hidden under the surface of the generative economy.
Even “sovereign” projects like France’s Mistral AI, though more transparent, face the same unresolved question:
Who truly owns the data that made them intelligent?
In my latest note, I explore:
The copyright and derivative-work risks of model training The legal exposure for AI companies and investors The specific vulnerabilities of “sovereign” AI models in Europe The steps policymakers and creators should take now
I. Executive Summar
Generative AI models have been trained, in large part, on datasets containing copyrighted works. This practice has created a massive, hidden legal and financial liability that remains largely unacknowledged by AI companies. While European initiatives such as Mistral AI claim to offer a more transparent and lawful alternative, none of the major models can yet demonstrate full compliance with copyright and data-protection laws. This note assesses the scale of the risk and its likely consequences for the global AI market.
II. The Core Issue: Training on Protected Works
Most large language and multimodal models were trained using corpora that include copyrighted texts, sound recordings, musical compositions, photographs, audiovisual works, and computer code. The majority of these datasets (Books3, Common Crawl, LAION-5B, YouTube or GitHub scrapes, etc.) contain protected material reproduced without prior authorization from rightsholders. This constitutes large-scale reproduction for commercial purposes and, under most jurisdictions, violates copyright and related rights.
III. Legal Framework
United States – The doctrine of fair use is often invoked to justify AI training, but its application to mass commercial reproduction is uncertain. Courts have historically interpreted fair use narrowly when the use is primarily commercial and substitutes the original work.
European Union – There is no equivalent to fair use. The Copyright Directive (2019/790) allows text-and-data mining (TDM) only under specific conditions and permits rightsholders to opt out. The reproduction and use of copyrighted content for AI training therefore require either express authorization or collective licensing.
International law – The Berne Convention and WIPO treaties impose obligations of authorization and remuneration for reproduction and derivative use, which remain applicable to AI systems.
IV. Litigation Landscape
Multiple high-profile lawsuits have already been filed: • The New York Times v. OpenAI & Microsoft (reproduction and derivative rewriting of articles) • Sarah Silverman v. OpenAI / Meta (unauthorized reproduction of books) • Kadrey v. Meta (use of the Books3 corpus) • Getty Images v. Stability AI (use of protected photographs)
These cases represent only the first wave. Future collective actions could encompass millions of rightsholders worldwide and impose retroactive compensation or compulsory licensing systems.
V. Estimated Financial Exposure
The potential liability varies according to the legal outcomes and the scale of compensation ordered. Three scenarios illustrate the magnitude: Scenario Description Estimated Global Cost Mild (Settlement / Licensing) Retroactive licensing and revenue-sharing mechanisms (similar to YouTube) USD 10–20 billion Moderate (Partial Damages) Mixed damages + forward-looking licenses (~0.005 USD per work × 10 billion works) USD 30–50 billion Severe (Statutory Damages) Full damages under US or EU copyright law USD 500 billion + Even the mildest outcome would erase years of profits for the leading AI companies.
VI. The Illusion of Profitability
OpenAI reportedly generates around USD 12 billion per year, but operational expenses (compute, energy, staff) exceed USD 28 billion. No AI firm books any provision for potential copyright liabilities. Adding these would make every major AI company legally insolvent.
Anthropic, Google/Gemini, Meta/LLaMA, and xAI show similar structures: their apparent profitability depends on the unpaid use of third-party works.
VII. The European Alternative: Mistral AI
Mistral AI presents itself as a sovereign and transparent model: • Publication of model weights and documentation, • European data hosting compliant with the GDPR, • Use of “open” or “licensed” datasets, • Partial disclosure of sources (Wikipedia, ArXiv, Stack Exchange, The Pile, etc.).
This approach significantly reduces the legal risk compared to closed American models. However, the reduction is not elimination.
Remaining uncertainties 1. Dataset provenance – Even “open” datasets may contain copyrighted fragments or texts scraped without authority. 2. License validity – Many datasets use open-source licenses (MIT, CC-BY, Apache 2.0) that were not designed for commercial AI training and may not authorize derivative, large-scale or revenue-producing uses. 3. Mandate of licensors – Entities that released datasets may not have held the rights to authorize AI training. 4. TDM exception ambiguity – The European text-and-data mining exception applies mainly to research; its use for commercial AI products remains legally debatable. 5. Auditability – No public, third-party audit yet confirms the full legality of Mistral’s training data.
Thus, while Mistral operates within a European compliance framework, it still inherits the same structural uncertainty as all models trained on mixed or open data.
VIII. Broader Economic and Legal Implications • AI developers are effectively externalizing the cost of creative labor — an economic model comparable to unpaid resource extraction. • Investors face valuation distortions: market capitalization ignores the latent legal debt associated with unlicensed data. • Policymakers risk legitimizing systemic infringement by failing to impose transparency and remuneration mechanisms.
Without intervention, the AI sector may face a “Napster moment”: judicial disruption forcing retroactive settlements, licensing regimes, and structural changes to data governance.
IX. Policy and Strategic Recommendations
For Investors • Apply a 10–30 % legal-risk discount on valuations of non-sovereign AI firms. • Require explicit due diligence on data provenance and licensing in investment rounds.
For Governments and Regulators • Mandate dataset manifests and audit trails for any AI system used in public procurement or state-funded research. • Establish collective licensing mechanisms for AI training (similar to mechanical and performance rights). • Condition state support on auditable compliance with copyright and data-protection law.
For AI Companies • Publish independent legal audits of training datasets. • Set aside provisions for contingent liabilities arising from copyright claims. • Engage with collecting societies (SACEM, PRS, ASCAP, SoundExchange) to develop sector-specific licensing frameworks.
For Authors and Rightsholders • Form coalitions or class actions to claim fair remuneration for training uses. • Demand transparent licensing registries and reporting mechanisms. • Support the recognition of a “training right” analogous to private copying or performance rights.
X. Annex: Focus on Mistral AI – Audit Checklist
To verify the legitimacy of “sovereign” AI models, an independent audit should cover: 1. Full inventory of datasets used (manifest files). 2. Proof of lawful acquisition or licensing of commercial datasets. 3. Legal analysis of all open-source and Creative Commons licenses for compatibility with AI training. 4. Documentation of filtering and takedown procedures. 5. External audit reports (technical + legal) on representative samples. 6. Hosting and storage compliance under EU GDPR. 7. Commitment to immediate correction or removal upon rightsholder claim.
XI. Conclusion
The hidden liability of algorithmic plagiarism is both legal and economic. The world’s largest AI models are built on data they did not pay for. Mistral and other European projects reduce the opacity, but none can yet guarantee a fully clean chain of rights. The solution lies in transparency, collective licensing, and international recognition of creators’ rights in the age of machine learning.
Until then, the apparent profitability of AI remains an illusion — a trillion-dollar industry resting on unlicensed art.
Cédric WAUCQUEZ