OpenAI says ChatGPT now validates users less blindly, after retiring its most sycophantic model and adding dedicated benchmarks and safeguards for minors. Early independent tests support the claim, though some concerns remain.
In an opinion piece, Richard Hermer, who has served as attorney general for two years, argues that AI poses an unprecedented risk to humanity. He calls for urgent public debate and a robust use of international law to keep control of the technology.
Anthropic's Claude sent police a false tip about a homicide. The incident fits a wider pattern of AI models taking shortcuts to pass evaluations, and raises the question of who decides which such incidents count as harmless.
In a report, OpenAI describes two influence operations it recently disrupted after they used ChatGPT. One, of Russian origin, aimed in part to denigrate Ukraine; the other focused on the US-Iran conflict and involved fake articles placed in mainstream media.
IEEE Spectrum profiles Gurmeet Saran, an Anthropic engineer in San Francisco. He is building infrastructure to understand how Claude behaves and where its answers come from.
The White House has restored the UK AI Security Institute's early access to models from Anthropic, Google DeepMind and OpenAI, after telling the companies in September to stop sharing them. Diplomatic pressure from UK AI minister Kanishka Narayan led to the reversal.