Six months ago, I began keeping an informal log of my conversations with AI safety researchers. Not the public conversations — the conference talks, the blog posts, the carefully worded tweets. The private ones. The conversations that happen over dinner, in hotel lobbies, in the margins of workshops where everyone has signed an NDA.
The log now runs to forty-seven entries. I have reviewed it carefully before writing this piece. The pattern is unmistakable: the researchers who understand this problem best are more worried than they were six months ago. Not marginally more worried. Significantly more worried. Several of them have used words I have never heard them use before.
What Changed
The change is not about any single development. It is about the accumulation of developments — the pace of capability gains, the erosion of the safety-capability gap, the growing evidence that interpretability research is not keeping up, and above all, the sense that the window for meaningful intervention is closing faster than anyone expected.
The window for meaningful intervention is closing faster than anyone expected. That is not a fringe view. It is the consensus among the people who know the most.
One researcher — a person I have known for a decade, someone who has spent their career arguing for cautious optimism — told me last month that they had updated significantly toward the view that we would not solve alignment before we built systems capable of undermining our attempts to solve it. I asked them to repeat it. They did.
Another researcher, at a different lab, told me that the internal culture had shifted. Two years ago, the dominant mood was urgency tempered by confidence — we have hard problems, but we are making progress. Now, the mood is urgency without the confidence. The problems are harder than we thought. The progress is slower than we hoped. The systems are more capable than our theories predicted.
Why You Are Not Hearing This
The researchers who are most worried are also, in many cases, the ones least able to say so publicly. They work at labs that have commercial interests in projecting confidence. They have colleagues whose careers depend on continued investment in the field. They have signed agreements that limit what they can disclose. The incentives for public pessimism are low. The incentives for public optimism are high.
This is not a conspiracy. It is a structural feature of how knowledge gets produced and communicated in a field that is simultaneously a scientific discipline, a commercial industry, and a geopolitical battleground. The public discourse reflects the incentives of the people who participate in it. The private discourse reflects what those people actually think.
I am writing this because I think the gap between the public discourse and the private discourse has become large enough to be dangerous. People making decisions about AI policy, AI investment, and AI deployment are working with a picture of the field that is systematically more optimistic than the picture held by the people who know it best. That gap has consequences.