The False Promise of Safe AI - The Atlantic

The Atlantic has dug up a line written by OpenAI researchers three months before ChatGPT launched, and it is a beauty: "there is currently no known indefinitely scalable solution to the alignment problem." Note the load-bearing adverb. Not no solution — no *indefinitely scalable* one, which is the same reassurance a landlord gives about a boiler that is fine *for now*. So the paperwork said the thing was unsolved, and the thing shipped anyway, because a caveat in a research memo is not a brake, it's a footnote with better formatting.

Four years on, the problem is still unsolved and — this is the part to savour — there is no consensus about what the problem actually is. Four years in which whole data centres have been poured into concrete and national grids quietly re-tuned. The industry can ship a model that writes your novel, sits your exams and, God help us, offers to date you, and still cannot settle what the alignment is meant to align *to*. Every launch deck meanwhile arrives with a safety section formatted like a car insurance policy: dense, unread, and technically sufficient. In current usage, alignment appears to mean the model not saying anything rude on the internet.

We were here a fortnight ago, when Greg Brockman urged AI safeguards while funding the political group opposing them — the now-standard choreography of a man applauding the fire brigade while paying the arsonist's retainer. The Atlantic's service is to hold up the receipts from before the beginning: the people closest to the work said there was no scalable fix, and the world got the product regardless. The alignment problem isn't unsolved because nobody is trying. It's unsolved because solving it first would have delayed the launch, and the launch was the point.