A warning from a recent case
The privacy tension is sharpened by the mass shooting in Tumbler Ridge, Canada, in February. The attacker had actively used ChatGPT while planning the attack.
OpenAI employees knew about the conversations before the shooting: the system had flagged them for human review months beforehand. The company’s leaders tried to hide ChatGPT’s role in preparing the attack and later played it down.
That history makes Anthropic’s decision look less like an isolated privacy incident and more like a test of what companies do with troubling signals. Anthropic has also faced serious problems involving real-world harm from its products, but appears to be trying to prevent a similar case.
The privacy question
Heller’s apparent belief that Claude was a diary points to a gap between how users may experience chatbots and how companies may handle their records. The account does not say what Anthropic reviewed, how it decided the threat warranted a report, or what safeguards governed that decision.
I think those omissions matter as much as the report itself. Companies are being asked to prevent harm, but the public has little basis here for judging where monitoring begins, how threats are assessed, or how users are made aware that conversations may be reviewed.
There are now so many chatbots on the market that preventing the next AI-linked tragedy may already be too late. The harder challenge is building credible ways to respond to danger without treating every private conversation as presumptively public.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X