Safety claims are becoming legal claims (Anthropic and AP)
Two AI safety stories landed in the same week. Read together, they show where accountability is heading.
Eight US state attorneys general have sued OpenAI, alleging the company deceived consumers about ChatGPT's safety, particularly for children. The claims are untested. But note the legal theory: not that the model is dangerous, but that what the company said about it was misleading. Safety marketing has become consumer-protection exposure.
Separately, Anthropic announced it has given a team from METR, a nonprofit model evaluator, access close to its own risk teams: company hardware, office space and internal tools. Six METR staff are onboarded. METR will publish its risk assessments, and Anthropic cannot veto publication. It can require redactions, but only in predefined categories, and METR is free to say publicly whether it thinks the redactions were justified.
The structure underneath is a tiered model. Narrow investigations sit at one level; publishing without approval sits at the next. Anthropic says it wants several evaluators at the deepest tier over time, each publishing its own analysis, and SecureBio is already reviewing the biosecurity chapter of its risk reports. The announcement follows Anthropic signing the White House accord on superintelligence, under which labs pledged to work with independent evaluators.
One caveat matters: Anthropic still decides what falls into the redaction categories. So the arrangement is a large step, not a finished one. Whether it proves meaningful depends on what METR eventually publishes.
As a former lawyer, I read these as the same development. When a company makes claims about safety, someone eventually tests them: an independent evaluator or a court. The first is cheaper, and you get to choose the timing.
This reaches well beyond the AI labs. Any organisation deploying AI to customers is making implied promises about accuracy, fairness and care. Those promises sit in your terms, your marketing and your sales decks right now.
Who has checked them against what the system actually does?
#RiskManagement #Regulation #AusBiz
Sources: Anthropic, 15 October 2026: https://lnkd.in/gtGxVBuv ; AP on the eight-state suits, 15 October 2026: https://lnkd.in/g-Ujp7fB