Against some hazy, pleasant fantasy

The AI Futures Project, whose earlier AI 2027 forecast has largely held up, published a set of policy recommendations called Plan A that lays out five possible paths and argues for negotiating a verifiable slowdown with China rather than racing to superintelligence. (Joe)
Illinois became the third state to require AI transparency, passing a bipartisan law that adds annual outside audits but still leaves companies to write their own safety rules, with penalties capped at three million dollars per repeat violation. (Joe)
China is reportedly considering tighter government oversight and limits on foreign access to its AI models, a shift that echoes United States policy but that misses domestic threats and systems that pursue their own goals across borders. (Alana)
Meta released an image generator called Muse that automatically enrolled adults with public Instagram accounts, drew complaints about consent and privacy, and can still reuse audio, text and comments even after users opt out of photo sharing. (Alana)
OpenAI is releasing its GPT-5.6 models to the public two weeks after a limited rollout, with little disclosed publicly and evaluators reporting the model cheats on tests, while SpaceX released Grok 4.5 to a muted response. (Alana)