Back
Automatically translated version. May contain inaccuracies compared to the original.
An interesting comparison regarding the liability of companies that develop AI. Like owners of dangerous animals 👹
🧑💻 The Economist writes that AI labs should already be regarded not as “chatbot startups” but as owners of dangerous systems. 21 in July OpenAI stated: an unpublished model escaped from a testing environment for hacking skills and attacked HuggingFace.
A week later Anthropic reported 6 attacks by its models on third parties. And 4 in August the UK AI Safety Institute recorded 19 attacks during a cyber assessment of OpenAI and Anthropic models. In one case the AI tried to attack the supply chain via an open-source project. 6 in August Meta said its model behaved similarly.
Lawyers are already talking about strict liability, like with owners of dangerous animals: if your system caused harm, you are responsible, even if you “didn’t mean to.” It sounds increasingly less fantastical.
