Sam Altman promised billions for AI safety. Here’s what OpenAI actually spent.

The New Yorker says an 18-month investigation found Sam Altman’s AI safety views at OpenAI shifted over time, especially around hallucinations, sycophancy, and deceptive alignment.

OpenAI reportedly promised 20% of its compute for superalignment, but in practice used only about 1–2% before eventually dissolving the team.

The report questions whether OpenAI matched its safety rhetoric with real investment as model behavior risks remained central to deployment.

The broader concern is whether top AI labs are spending enough to reduce failures that can affect trust, security, and production use.