The Doom Predictions, Graded
A decade of AI executives warning the world about their own product, checked one by one against what actually happened. Most of it is unfalsifiable marketing. Three predictions were specific enough to test — and the newest grade just came in, from Anthropic's own numbers.
Explain Like I'm 5 — the whole thing in plain language.
AI company bosses have warned for years that their own product could be dangerous. Almost none of those warnings can be checked, because they name no date and no mechanism. That makes them free advertising: a scary claim that can never be proven wrong costs nothing to keep repeating.
Musk has said AI is civilization's biggest risk since 2017, and later gave it a rough percentage chance of going badly, a number borrowed from another researcher, with no way to test it. Altman told Congress in 2023 he wants his industry regulated. Hundreds of researchers and executives signed a one-line statement comparing AI risk to nuclear war. None of that is checkable. It's a position, not a prediction.
Only three doom claims from the last decade named something specific enough to grade. OpenAI once said its GPT-2 text generator was too dangerous to release in full, then released it months later anyway. Nothing bad happened: wrong. In 2025, Anthropic's CEO predicted AI would wipe out half of entry-level office jobs within five years; early data leans his way, but it's too soon to call. And back in 2023, he predicted AI could help build bioweapons within two to three years. In September 2026, Anthropic's own safety report admitted, in writing, that newer models have crossed exactly that line.
The lesson: a scary claim with no date is just marketing. A scary claim with a real date is a prediction, and this year one of them started coming true.