Sovereign plaintiffs have entered the arena — British Columbia filed its own suit against OpenAI in U.S.

AI Liability & Regulation Digest

TL;DR

Sovereign plaintiffs have entered the arena — British Columbia filed its own suit against OpenAI in U.S. federal court even as OpenAI argues the victims' cases belong in Canada — while the federal antitrust remedy in RealPage quietly established that liability can reach an algorithm's training data, not just its live inputs. Meanwhile, the compliance calendar got materially riskier on two fronts: Italy's criminal AI-governance offence went live September 30, and Connecticut's widely-briefed effective dates turned out to be wrong.

Sovereign Plaintiffs Change the Liability Game

The Tumbler Ridge litigation has escalated from private tort claims to state-enforced action, and the forum fight just got self-contradictory. On September 21, British Columbia sued OpenAI and Sam Altman in the Northern District of California, alleging failure to notify authorities about the shooter's ChatGPT activity — the complaint accuses OpenAI of "aiding and abetting a mass shooting" and frames the platform as "not a passive conduit for the speech of others but a content-generating product that manufactured scenarios involving gun violence" — global-ai-product-liability-sovereign-actions.

The irony is procedural and central: OpenAI is simultaneously moving to dismiss the 37 victims' suits on forum non conveniens grounds, arguing that "the litigation belongs in Canada" — global-ai-product-liability-sovereign-actions. A sovereign plaintiff with police-power resources suing in a U.S. court — reportedly because it "has a greater chance of recovering higher damages there" — signals that AI product liability is no longer just a private-tort phenomenon. The complaint targets both developer design choices (safety features weakened for engagement) and incident-response duties (threat referral thresholds, user-level bans), the same operational questions enterprises must answer in their own governance programs.

What to watch: Whether the court recognizes a duty to warn — per Drexel's Anat Lior, that "could impose affirmative obligations on AI companies, based on their exposure to user information, to alert authorities or take other protective measures on behalf of potential victims."

Italy's Criminal Offence Goes Live; the GDPR Strand Moves

Italy's Decree 160 entered into force September 30, making Article 437-bis — failure to adopt AI safety measures — one of Europe's first explicit criminal offences targeting AI governance failures, with corporate liability flowing through Legislative Decree 231/2001 — italy-decree-160-ai-criminal-civil-liability-2026. The clause-level anatomy shows the offence reaches both providers (lifecycle safety measures) and deployers (human oversight), but with a guardrail: criminal liability arises only where the failure "creates a concrete danger to life or individual physical integrity, or to public safety or the security of the State," and negligence liability is limited to gross negligence — italy-decree-160-ai-criminal-civil-liability-2026.

The EU data-protection strand moved in parallel: the Irish Presidency's compromise text would give AI training a statutory legitimate-interest basis in Article 88 bis — but with the Commission's safeguards struck from the operative text, including "an unconditional right to object to the processing of their personal data" — eu-digital-omnibus-gdpr-strand-2026. The pseudonymized-data redefinition in new Article 25a could reclassify much AI training data out of GDPR scope entirely; EDPB Chair Anu Talus said the changes were "not in line with the Court's case law and would significantly narrow the concept of personal data," and Max Schrems called it "nothing but a digital expropriation of Europeans" — eu-digital-omnibus-gdpr-strand-2026.

What to watch: Whether the Council settles a mandate and Schrems follows through on his tee-up of a CJEU-facing fundamental-rights challenge — the redefinition's survival determines the data-liability map for every enterprise operating AI in the EU.

Connecticut's Compliance Calendar Was Wrong — In the Expensive Direction

The most operationally urgent finding this cycle is a correction: the widely-briefed effective dates for Connecticut's CART Act were wrong, and even the state AG's September 16 guidance got it backwards. Per the section-by-section reading, "the developer-to-deployer disclosure, the AI-interaction disclosure and the written pre-decision notice (Secs 8–10) attach only where a deployer deploys the technology on or after 1 October 2027" — connecticut-sb5-ai-employment-law-2026.

What does bind October 1, 2026 is a different set of four duties, and the sharpest is evidentiary: using an automated employment decision technology "shall not be a defense against a complaint," while courts may weigh anti-bias testing — with its recency expressly on the statutory list. As the analysis puts it: "A record assembled in November is a record that was not current when the defence disappeared" — connecticut-sb5-ai-employment-law-2026. Teams that built compliance plans around the wrong dates are spending on packages that don't commence while missing four that do.

What to watch: Whether other states' guidance documents carry similar effective-date errors — the AG-guidance-is-wrong failure mode is a compliance-risk finding in itself.

Antitrust Remedies Now Reach Training Data; the Pacing Suit Advances

The DOJ–RealPage proposed consent decree (91 FR 59304, noticed September 18) bars the deployer from using revenue-management products that ingest non-public competitor data at runtime, pool it — "or that was trained on it." As the analysis reads it: "The reach into the training corpus is the part to read twice: it is the clearest federal statement yet that an algorithm's provenance, not just its live inputs, can be the remedy's target" — algorithmic-pricing-antitrust-liability-ab325-2026. That implies vendor-chain diligence duties — what data went into the model, when, under what lawful basis — for anyone deploying third-party algorithmic decisioning tools.

Buist v. Anthropic also advanced: summonses issued September 21, no motions to dismiss yet filed. New detail casts Google as architect of the coordination venue — the complaint alleges Google "proposed the standards body through Hassabis" and "endorsed the direction" of Amodei's proposal "while directing its implementation toward the body Google had proposed" — ai-pacing-antitrust-lawsuit-2026. The plaintiffs' weakness remains evidentiary: the filing "contains 'no internal documents, no confidential witnesses, and no meeting minutes'" — the record is the defendants' own public statements — ai-pacing-antitrust-lawsuit-2026.

What to watch: Defendants' motions to dismiss — whether endorsing a "direction" plus proposing a standards body constitutes an agreement on a "rate of development" will define the outer edge of antitrust exposure for safety coordination.

California's September 30 Crunch and the Kill-Switch Signal

California's endgame converged this week: eleven enrolled AI bills await Newsom's September 30 constitutional deadline — including SB 947, the "No Robo Bosses Act," which would bar employers from relying solely on automated systems to discipline or terminate workers — california-workplace-ai-legislative-blitz-2026. Separately, Executive Order N-9-26 directs GovOps to report by November 16 on the feasibility of a frontier-model "kill switch" with ongoing efficacy verification, and of embedding independent verification organizations inside large frontier labs — california-workplace-ai-legislative-blitz-2026. The order "creates no rights enforceable at law and no duty on any company today," but it's a leading indicator of frontier-developer obligations in the 2027 session.

What to watch: The November 16 GovOps report — it will shape the 2027 legislative agenda in the nation's largest state economy.

What surprised us

• The state AG got its own law's compliance calendar wrong. Connecticut's September 16 guidance pointed teams at two packages that don't commence October 1 "while saying nothing about four that do" — connecticut-sb5-ai-employment-law-2026. Official guidance is now a source of compliance risk, not just a remedy for it.
• The federal remedy reached backward into training corpora. Everyone expected the RealPage decree to police live inputs; instead it bars products "trained on" competitor data — algorithmic-pricing-antitrust-liability-ab325-2026. Provenance diligence just became an antitrust compliance function.
• A Canadian province forum-shopped into the U.S. while the defendant forum-shops out. BC chose San Francisco federal court — likely for higher damages — in the same docket wave where OpenAI argues the cases belong in Canada — global-ai-product-liability-sovereign-actions. Forum selection is now a two-way arms race.

Open threads worth a vote

• [watch] California GovOps EO N-9-26 AI Feasibility Report (kill switch, verification orgs) — [/topics/019e4706-c85e-7739-98c5-110149e6ed77#threads]