AI Liability & Regulation Digest
TL;DR
The liability map is being redrawn on two fronts at once: Italy just became the first country to convert EU AI Act obligations into criminal and civil liability — with prison terms that reach deployers, not just developers — while U.S. courts are simultaneously narrowing some developer exposure (the Ninth Circuit's DMCA ruling) and widening other fronts (Florida's AG suit proceeding before a county jury, and a novel antitrust suit targeting AI safety coordination itself). The through-line: procedural and venue fights are deciding more right now than any merits ruling.
Forum Fights Are the New Front Line
Venue is hardening into the decisive battleground for AI harm litigation, and developers just lost the first round. Judge Aileen Cannon granted Florida's motion to remand its consumer-protection suit against OpenAI back to Highlands County state court, warning that broad removal "could 'herald a potentially enormous shift of traditionally state cases into federal courts'" — florida-ag-v-openai-lawsuit-2026. Meanwhile, OpenAI is running the opposite play in the Tumbler Ridge wrongful-death wave, filing identical forum non conveniens motions arguing that "California is not the most appropriate forum for claims by British Columbians attacked by a British Columbian in British Columbia" — us-ai-wrongful-death-lawsuits-developer-liability-2026
.
The pattern matters because forum determines jury pool, discovery posture, and governing standards of care — and the Florida ruling hands other state AGs a template: plead federal standards like COPPA as evidence of unfairness, and keep your case in state court. For enterprises deploying consumer-facing AI, venue strategy is now a compliance input, not a litigation afterthought.
What to watch: The Northern District of California's ruling on OpenAI's forum non conveniens motions — the last open venue question from the previous cycle's digest, and one that will shape the entire cross-border AI-harm litigation wave.
Courts Narrow Novel Theories, Leave the Big Questions Open
The first appellate ruling in the AI-training copyright wave cut against plaintiffs — but only on the thinnest of the theories. The Ninth Circuit upheld dismissal of the DMCA claims in Doe v. GitHub, finding "Copilot and Codex do not remove copyright information from existing works, but instead generate new works that never contained such information" — ai-copyright-liability-cases-2026-trends. Crucially, the panel expressly reserved whether output similarity could support traditional infringement, and the open-source licensing claims below are still live.
That fits the emerging pattern: courts are pruning creative statutory theories while leaving fair use and substantial-similarity infringement — the questions with real money attached — untouched. Enterprises shipping code-generation features should treat open-source license attribution and output-similarity indemnification as the two live contract terms to negotiate now.
What to watch: Whether the open-source licensing claims survive against GitHub, since they remain the only live theory in the case and a preview of vendor indemnification fights.
Italy Turns the EU AI Act Into Criminal and Civil Liability
Italy's Legislative Decree No. 160, published September 15, is the first national framework to attach prison terms and entity liability to AI Act non-compliance — and it reaches deployers, not just providers. New Article 437-bis punishes anyone who "fails to adopt human oversight measures" for high-risk systems with one to five years' imprisonment where the omission endangers life or safety; as the analysis notes, "Deployers, not only providers, sit inside the offence" — italy-decree-160-ai-criminal-civil-liability-2026.
The civil chapter may matter more commercially: courts can order production of AI Act logs and risk-management documentation, with non-compliance treated as an admission of the facts alleged; breach of AI Act obligations triggers a presumption of causation; and claimants can sue the operator's insurer directly. The timing is the sting — national liability regimes are arriving ahead of both the harmonised high-risk compliance deadline (now December 2, 2027, per the enacted Digital Omnibus: "stand-alone high-risk systems listed in Annex III 'now become subject to the regime on 2 December 2027, instead of 2 August 2026'" — eu-ai-act-omnibus-may-2026) and the December 9, 2026 Product Liability Directive transposition deadline. The Article 9/11/12/14 documentation you'd build for 2027 is becoming discoverable litigation evidence today.
What to watch: Whether other member states copy the Italian model as the PLD transposition deadline hits in December — the deferral runway is not a liability runway.
California Signs the Audit Mandate; the "Human Override" Defense Hangs in the Balance
California's slate moved from pending to partially signed, with Governor Newsom signing SB 1119 ("Adam's Law") on September 10 — "the only law in the US requiring operators of such products to (1) perform and document a 'comprehensive risk assessment' with respect to child users, and (2) submit to 'independent child safety audits'" — california-workplace-ai-legislative-blitz-2026. Still on his desk: SB 947 ("No Robo Bosses Act"), which would make human corroboration of automated employment decisions a statutory requirement, and AB 2575, which would prevent clinical AI developers and deployers "from pointing to such choices as a defense in lawsuits alleging harm from CDSS output."
AB 2575 deserves particular attention because it targets the exact escape hatch — "a human reviewed and overrode the AI" — that most clinical AI deployments rely on for liability cover. Newsom vetoed the prior version of SB 947 as overbroad in October 2025, so the veto watch from last cycle's digest extends, but the signed audit infrastructure (SB 813/AB 1405's third-party auditor capacity-building) means the enforcement supply chain is being built regardless.
What to watch: The September 30, 2026 signing deadline on SB 947 and AB 2575 — the two bills that would most directly reshape developer/deployer liability allocation in the U.S.
What surprised us
- Safety coordination is now an antitrust theory. Buist v. Anthropic alleges that Anthropic, OpenAI, SpaceXAI, and Google violated Sherman Act Section 1 by agreeing to slow AI development after Amodei's "We Must Pace the Frontier" essay, with Altman, Musk, and Hassabis endorsing it the same day — ai-pacing-antitrust-lawsuit-2026
. The compliance world has spent years encouraging shared benchmarks and joint safety frameworks; this suit argues those same mechanisms are an output-restraining cartel. That's a genuinely disorienting inversion.
- A Trump-appointed judge handed plaintiffs a venue win. Cannon's remand rejected both removal pathways and declined to sanction Florida, but the practical effect is that the first state AG sovereign AI suit will be heard by a Highlands County jury — florida-ag-v-openai-lawsuit-2026
. Strict liability and public nuisance claims against a chatbot developer, in state court, are now the live docket.
- The EU gave with one hand while Italy took with the other. The Digital Omnibus deferral bought 16 months on Annex III high-risk obligations — eu-ai-act-omnibus-may-2026
— but Italy's decree shows member states won't wait to attach consequences to the underlying obligations. The runway is for building, not for breathing.