Law and society - GeneseAI

21.1Copyright and intellectual property

Diagram21.1. The two fronts of copyright. U.S. courts tend to regard training as a "transformative" use, but the question of whether the content produced unlawfully reproduces existing works, and who answers for it, remains entirely open.

Under the hood

In 2025-2026, U.S. judges held that training a model could qualify as fair use, comparing an AI's learning to that of a human reader. But the question of pirated works remains weighty: a major lab agreed to pay around 1.5 billion dollars to settle litigation tied to the illegal downloading of millions of books. Several actions are underway (publishers against Meta, a dozen cases consolidated against OpenAI). On the image side, major film studios (including Disney and Universal) are suing an image generator, while other players are signing licensing deals: the emergence of a licensing market makes a fair use defense harder to sustain. Finally, on the protection of works produced by an AI, the U.S. Supreme Court, in March 2026, declined to hear the landmark case on the matter (the Thaler case): it thereby leaves in force the 2025 ruling by a federal appeals court, which held that a work generated by a machine without human creative contribution is not protectable.

21.2Deepfakes, disinformation and democracy

In context

Mass persuasion

Distinct from disinformation (which rests on the false), another concern is mounting: persuasion itself, that is, the ability to change minds, even with truthful arguments. A preregistered study (Salvi et al., Nature Human Behaviour, 2025) had around 900 people debate either a human or GPT-4 on social issues. A striking result: when the model had access to a few sociodemographic details about its interlocutor, it proved more persuasive than the human in roughly 64% of the unequal debates; without personalization, it was on a par. The decisive advantage is not so much the quality of an argument as the ability to personalize instantly and at vast scale: to tailor the message to each profile, by the millions, in real time. One can gauge the risk in marketing, political campaigns or social networks, where such automated microtargeting could influence opinion on a massive and discreet scale. The authors themselves call on platforms and regulators to monitor, and where necessary to regulate, the use of these models in sensitive persuasive contexts. It is a societal issue in its own right, at the crossroads of psychology (chapter 19) and democracy.

21.3Privacy and data protection

In context

Differential privacy

How can sensitive data (health records, a service's usage logs) be exploited without exposing the individuals who make it up? A rigorous answer, now a standard, is differential privacy. The idea: add to the results a carefully calibrated dose of random noise, so that the aggregate statistics remain reliable, but it becomes mathematically impossible to tell whether a given person was or was not in the dataset. One can thus publish trends, or even train a model, while formally bounding what can be relearned about an individual. The technique, used by major players (statistical institutes, phone makers collecting usage without tracking each person), comes at a cost: the more you protect, the more you degrade accuracy, hence a trade-off to settle. Combined with federated learning and encryption (chapter 11), it is part of the toolkit of a "privacy-respecting by design" AI, an alternative to centralizing data in the clear.

21.4Surveillance, bias and discrimination

In context

Linguistic diversity and the Global South

The benefits of AI are not distributed equally across languages. The best models are trained mainly on English and a handful of data-rich languages; they are markedly less capable in the thousands of low-resource languages, often spoken in countries of the Global South, which risks widening a digital divide (the populations concerned also being less well served by infrastructure and by the cost of the models). Efforts exist to address this: projects to build models for African or Indian languages, open and community-driven datasets, linguistic sovereignty initiatives. The stakes are not only technical: a language poorly served by AI sees its digital use recede, which can accelerate its erasure. The linguistic diversity of AI is therefore also a cultural and an equity issue.

21.5Liability, justice and the positive side

In plain terms

Three issues round out the picture.

Liability. When an AI causes harm (a faulty diagnosis, an autonomous-car accident, ruinous financial advice), who is responsible? The user, the deployer, the designer? The idea of granting "legal personhood" to an AI is largely set aside; the trend is rather toward product liability regimes and an obligation of human oversight. The European Union has, moreover, modernized its product liability regime to explicitly include software and AI systems.
Predictive justice. The use of algorithms to assess a risk of reoffending or to guide judicial decisions has drawn sharp criticism: some tools have proved biased to the detriment of certain populations, raising the question of equality before the law.
The power of platforms. AI now moderates billions of pieces of content: a considerable power over free expression and information, concentrated in a few hands, with errors at scale. The protection of children online (inappropriate content, AI companions, chapter 19) is a particularly sensitive case of this.

In context

The "right to an explanation."

When an automated decision affects a person (a credit refusal, a candidate screening, a benefit denied), can they demand to understand how it was made? This is the issue of the right to an explanation, sketched out in European law (the GDPR frames decisions "based solely on automated processing" and opens a right to human intervention and to contest). The principle is powerful, but it runs up against the opacity of models (the "black box," chapter 2): explaining the output of a network of several billion parameters in intelligible terms is an unsolved technical problem, which connects to research in interpretability (chapter 24) and explainable AI. Hence pragmatic approaches: providing the main factors in a decision, guaranteeing a human recourse, documenting the system. The stakes go beyond user comfort: without explanation or recourse, there is neither possible challenge nor effective accountability, and the algorithmic decision escapes all democratic control.

Key takeaways (chapter 21)

Copyright: training tends toward fair use in the United States, but piracy is penalized and liability for generated content remains open; a licensing market is emerging.
Deepfakes: fraud, disinformation, a threat to democracy; the deeper danger of the "liar's dividend"; partial countermeasures (detection, traceability, mandatory labeling in the EU).
Privacy: AI can infer sensitive data; the GDPR imposes strong constraints in Europe, with the challenge of articulating it with the AI regulation.
Surveillance and bias: uses varying by region, a risk of algorithmic discrimination hitting the most vulnerable.
Liability (no legal personhood for AI), predictive justice (bias), the power of platforms; but also a major positive side: accessibility for people with disabilities.

From civil law to the most sensitive sovereign domain: chapter 22 deals with defense and international security.

21.1Copyright and intellectual property#

21.2Deepfakes, disinformation and democracy#

21.3Privacy and data protection#

21.4Surveillance, bias and discrimination#

21.5Liability, justice and the positive side#

Key takeaways (chapter 21)

21.1Copyright and intellectual property

21.2Deepfakes, disinformation and democracy

21.3Privacy and data protection

21.4Surveillance, bias and discrimination

21.5Liability, justice and the positive side