AI harms in court — litigation, rulings and settlements

When a new technology causes harm, the first record is usually the headline. The second, much slower and much quieter, is the court file. This tracker deals with the second kind: the hundred cases in which harm attributed to artificial intelligence stopped being a controversy and became a lawsuit, a sanction, a formal complaint to a regulator, or a settlement with an identifiable outcome. The distinction matters. There are thousands of AI incidents; only those a legal system agreed to examine make it in here.

The result is a dataset spanning from 2012 —when a Medicaid algorithm in Idaho cut disability benefits by up to 42% and the ACLU took it to court— to 2026. But the bulk of the phenomenon is recent: more than half the cases (51 of 100) cluster in 2024 and 2025. It is not that there was no harm before, but that there was barely a legal avenue to name it.

Seven fronts, not one

The first finding is that 'AI harm' is not one category but seven. The tracker splits the hundred cases across deepfakes (23), facial recognition and biometrics (23), automated decision-making (18), bias and discrimination (12), chatbots (9), defamation (8) and autonomous vehicles (7). That dead heat at the top —deepfakes and biometrics— captures two very different fears: that a machine will fabricate a false reality about you, and that it will misidentify you in the real one.

Within deepfakes, the dominant subtype is non-consensual sexual imagery: adding adult victims, minors and public figures together, it outweighs any other use, including voice-cloning fraud. The Almendralejo case in Spain —fifteen minors convicted of spreading AI-generated fake nudes of female classmates— and the Seoul National University case in South Korea —ten years in prison for 1,852 synthetic sexual images— mark the two extremes of a single trend: criminal justice is responding, but the severity of sentences varies enormously across jurisdictions.

Automated decision-making, the third category, is the most institutional of all. These are algorithms the state or insurers use to allocate or deny resources: the Netherlands' SyRI welfare-fraud detection system, ruled unlawful by a Hague court in 2020; the unemployment profiling that Poland's Constitutional Tribunal struck down in 2018; the 'suspicion score' algorithm of France's CNAF, challenged in 2024 by fifteen organisations. Here the harm is not an accident: it is the design working as intended, on the population least able to appeal.

Who litigates what, and where

The geographic split is not neutral: it reveals which legal tool dominates in each region. In the United States, the civil lawsuit rules —class action, product liability, defamation—, as in the $329 million verdict against Tesla over an Autopilot failure or the Clearview AI national settlement valued at $51.75 million. In Europe, the data regulator is the protagonist: the GDPR has let the authorities of Italy, the Netherlands, France, Greece and Sweden fine Clearview AI more than €110 million combined, though the company has ignored the payments. In Asia, the weight falls on criminal prosecution: prison sentences for deepfakes in South Korea, arrests for voice-cloning scams in Hong Kong.

That diversity of avenues explains why the 'status' or outcome field is the most revealing data point in the tracker. Fifteen cases remain ongoing; the rest divide their fate among final fines, compensation settlements, prison sentences, systems declared unlawful and —a fact worth not forgetting— dismissals. Walters v. OpenAI, the first defamation suit over a chatbot-generated falsehood, ended with a Georgia judge ruling for the company in 2025: the product's disclaimers and the absence of malice were enough to defeat the claim. Europe's first major fine against a generative AI, the Italian Garante's €15 million against OpenAI, was annulled by the Court of Rome in 2026. The record, in other words, is not a string of victories against the technology; it is a contested field.

The dataset itself carries a bias worth declaring: it depends on a legal system being able and willing to record the harm, and on that record being public and verifiable. That under-represents countries without such infrastructure and over-represents the United States and Europe, where litigation is more common and media coverage denser. The absence of cases in a region does not mean the absence of harm.

Why this tracker exists

This tracker is the third leg of an ecosystem. The first documents the fines regulators impose on digital platforms; the second, governments' public spending on AI. Between the two lies a gap: the human and judicial cost, what happens to specific people when a system fails and someone decides to fight it. Together the three close a triangle —what the state punishes, what the state spends, and what the state, through its courts, repairs or refuses to repair.

That is why what is measured here is not the uproar but the procedure. A viral deepfake of a celebrity that generates only online outrage does not qualify; one that ends up in a courtroom does. The gap between the amount of harm that gets reported and the fraction that reaches a legal record is, in itself, the story this tracker aims to make visible.