Law & AI: Regulation, Liability, and the Courts

Eight banned uses including social scoring, untargeted facial-recognition scraping, and most real-time remote biometric ID in public. Plus AI-literacy duties.

General-purpose AI model providers face transparency, copyright-policy, and systemic-risk duties. Governance bodies stood up.

Conformity assessments, risk management, human oversight for AI in hiring, credit, education, and justice. General date Aug 2, 2026; product-embedded systems to 2027–2028.

Up to €35M or 7% of global turnover for prohibited use; €15M or 3% for other breaches — whichever is higher.

First comprehensive US state AI law (SB24-205). Effective date pushed to June 2026, then superseded by SB 26-189 — now effective Jan 1, 2027. A live example of regulatory churn.

Responsible AI Governance Act, effective Jan 1, 2026. Intent-based: bans specific harmful uses rather than EU-style risk tiers.

Voluntary federal framework (AI RMF 1.0, Jan 2023). The Generative AI Profile (Jul 2024) names 12 GenAI risk areas. Soft law, widely cited in contracts.

Federal action runs through executive orders and agency guidance, which shift with administrations — leaving states as the primary lawmakers.

Judge Alsup: training on legally bought books is "exceedingly transformative" fair use — but using pirated copies is not. Followed by a record $1.5B settlement over pirated works.

First major AI-training copyright ruling. Copying 2,243 Westlaw headnotes was infringing, not fair use — but expressly limited to non-generative AI.

Court denied most of OpenAI’s motion to dismiss (2025); core infringement claims proceed to discovery. The marquee fair-use trial still ahead.

High Court, Nov 2025: model weights are not a "copy" under UK law; Getty dropped its primary copyright claims mid-trial. Narrow trademark win only.

Thaler v. Perlmutter (D.C. Cir., 2025): the Copyright Act requires a human author. SCOTUS denied cert in 2026, settling the rule. The Copyright Office agrees.

The EU abandoned the AILD in 2025 for lack of consensus — leaving AI-as-a-service fault liability to 27 fragmented national regimes.

Revised PLD (adopted 2024) expressly covers software and AI as products, addressing post-sale updates and machine learning.

When an AI service causes harm, who pays — developer, deployer, or user? In the EU this now depends on which member state you are in.

79% of legal professionals reported using AI tools in 2025, up from 19% in 2023. Roughly half of Am Law 100 firms use Harvey.

The origin case: $5,000 Rule 11 sanction for ChatGPT-fabricated citations. Now the cautionary tale cited in standing orders nationwide.

Damien Charlotin’s database logged ~712 cases by Dec 2025 and ~1,227 by early 2026, with 5–6 new filings per day at peak.

~90% of proceedings end in warnings or struck material; fines, when imposed, run from a few hundred to several thousand dollars.

Wisconsin upheld COMPAS risk scores in sentencing but required warnings: scores rely on group data, were never validated locally, and cannot be determinative.

Found Black defendants falsely flagged high-risk at 44.9% vs 23.5% for white defendants — the study that put algorithmic bias on the legal map.

Robert Williams (Detroit, 2020) was the first known US wrongful arrest from a face-match; Detroit settled for $300,000. At least seven such cases are public.

Gives individuals the right not to be subject to solely-automated decisions with legal effect — the EU’s long-standing backstop, now reinforced by the AI Act.