Sam Altman: Warnings vs Promises

In May 2023 Sam Altman told a Senate subcommittee that AI could “go quite wrong”, asked for a licensing regime, and signed the statement putting AI extinction risk alongside pandemics and nuclear war. That was the high-water mark for his warnings, and it has not been matched since.

This analysis scores four years of his public statements on two axes: how large a capability claim each one makes, and how much of it is about harm from AI. Every statement sits on its actual date, so the gaps on the chart are true to elapsed time.

What it found

The gap between the two readings, capability minus danger, sits near zero for the statements up to the May 2023 hearing, and is positive almost everywhere after it. It crosses back only at governance venues: the January 2026 biosecurity town hall and the February 2026 India summit. The claim is not that Altman is wrong about where AI is going, but that his risk talk and his capability talk now answer to different pressures.

How it was scored

• Rubric first. Both axes are ordinal bands with fixed examples, written before the corpus was assembled.
• Randomised order. Statements were scored in random order, each score naming its band, with no revision afterwards.
• Selection, recorded. The 23 statements are the moments where his framing visibly changed, a deliberate choice that is written down rather than hidden.
• Robustness. The result survives leave-one-out and two bands of simulated scoring error. A funding-timeline explanation was tested and dropped: OpenAI raised money almost continuously, so it explained nothing.

The honest caveats

The scoring is one person’s reading, made knowing what the analysis was looking for. With only four statements before the split, the earlier period is description, not evidence. The test that would settle it is a second coder scoring the blank sheet in the repo cold; if the agreement is weak, the shape of these lines is the author’s rather than the record’s.