Europe's strict product liability reaches standalone software on 9 December

Sacks, who chairs the President's Council of Advisors on Science and Technology and was the White House AI and crypto czar until March ^[2], offered a hypothetical that Europe has already written down. If a model enables a serious cyberattack, he argued, the company faces product liability claims, and markets already punish systems that behave unpredictably ^[5]. From 9 December the EU's defectiveness test takes in cybersecurity vulnerabilities and self-learning behaviour that turns unsafe after release ^[12].

For whoever owns the release, the change is evidentiary. A claimant does not have to prove fault ^[11]. Courts can order the defendant to disclose evidence, and can presume a defect where the claimant faces excessive difficulty proving one ^[13].

Sacks was answering two chief executives ^[1]. The rule covers standalone software and operating systems alongside AI systems ^[11]. A great many teams sit inside it who are not building a frontier model ^[17]. European lawyers have been working through what that means since the summer, in driver assistance as much as in chatbots ^[16].

What he refused was the price attached to restraint. Companies asking to pace the frontier should stop pretending antitrust law must be suspended "so you can form a cartel", he wrote, and stop demanding an approval process that supersedes product liability ^[4]. Demanding a regulatory framework as the price of restraint, he wrote, would look like "blackmail" ^[6]. He also questioned the independence of METR, the evaluator Anthropic uses, saying it is intertwined with the company's investors and staff ^[7].

Europe's arrangement is different from the swap he describes. The Commission withdrew its proposed AI Liability Directive in February 2025 after deciding the revised product liability rules already covered AI software ^[10]. But the AI Act's systemic risk duties have bound frontier models since August 2025, and since 2 August the Commission can demand evaluations and fine 3% of global turnover ^[14]. Both regimes will operate on the same models at once ^[15].

The argument is playing out alongside a resignation. A pretraining researcher left Anthropic on 8 September saying the industry was gambling with people's lives ^[8], and the company's alignment science lead put the odds of AI causing human extinction within a decade above 10% ^[9]. Sacks put the responsibility back on the labs. "The easiest way not to build superintelligence is for you to agree not to build it," he wrote ^[3].

What sorts the exposure for a team shipping into the EU is whether what you shipped can change its behaviour after release, and whether such a change can be explained from records that exist somewhere other than a chat thread. Where behaviour changes and the explanation is undocumented, a court can order disclosure and presume a defect ^[13].