How the record works
Anyone can stamp a claim false. Here is the rule we follow, so you can hold us to it.
What gets on the record
A claim goes on the record only if it has a named person who said it in public, a date, a link to the original, and a deadline. If a claim is too vague to ever be settled, it does not go on the record — however interesting it is.
The test is written first
Every claim carries a settlement criterion — the specific thing that would have to happen for it to count as true. It is written when the claim goes on the record, published on the claim's page, and locked from that point.
This is the part that stops a verdict being hindsight. If we only decided what counted after seeing the outcome, we could fit the test to whatever happened and nobody could tell. If a criterion cannot be written, the claim cannot be settled, and it does not go on the record.
Criteria can be changed, but never quietly. Any change is logged below with the reason, alongside every other correction.
How we read a claim
We settle a claim as a reasonable reader would have understood it when it was made — not the most literal reading, and not the most generous one. If someone said “within six months,” the deadline is six months from the date they said it. If they gave a range, the deadline is the far end of it.
Hedges are part of the claim. When someone says “may”, “probably” or “my guess is”, that wording stays in. Tightening a hedge into a flat forecast makes a person look more certain, and more wrong, than they were.
The three verdicts
True — the claim happened by the deadline.
False — it did not.
Unclear — the claim turned out not to be settleable, or the evidence is genuinely contested. Unclear is a real outcome, not a failure. We would rather say unclear than force a claim into a stamp it doesn't fit.
What counts as evidence
Every verdict links to the evidence behind it. Where the evidence is partial or disputed, we say so on the record rather than in private.
How votes are counted
Voting takes no account and no email. One vote per browser per claim, and one per internet connection per claim. Clicking your own answer again withdraws it. The reader split stays hidden until you have voted, so that seeing the crowd cannot pull your answer towards it.
This is best effort, not an audit. Someone determined to vote repeatedly, across different connections, could do it. We would rather say that plainly than imply a rigour we do not have. Where a count came from a poll run elsewhere, the source is printed next to the number.
Reader votes never affect anyone's record. They are what people thought would happen, kept separately from what did.
Who decides
A verdict is settled by the editor, working from the criterion published when the claim went on the record. There is no panel vote and no algorithm. Saying so plainly is the point — you can weigh the judgement knowing how it was reached.
The record is maintained by a small group of volunteers. Anyone can send in a claim; what goes up is checked against the original first.
If you think we got it wrong
Send us the evidence, or write to hello@aionrecord.com. Verdicts can be reversed, and reversals are logged publicly with the date and the reason. Nothing on this site is edited silently.
What the expert labels mean
The line under each quote is our characterisation of that person's position, not a verdict on whether they were right. The quote is always there so you can disagree with our reading.
Corrections
Every change to a published verdict or entry is logged here, with the date and the reason.
Split a compound claim by Roman Yampolskiy — The entry combined an AGI date with a separate statement about occupations being replaced. Two claims in one row cannot be settled, because one half can be right while the other is wrong. It now covers the dated half only.
Settled the Altman agents claim as unclear — Its deadline passed in December 2025 but the entry was left open by an import error. The claim offered no measure of what would count as materially changing company output, so it settles as unclear.
Traced the Musk timeline entries to their original occasions — The 2029 step rested on a passing mention in secondary coverage. It is now dated to the July 2023 Twitter Spaces call announcing xAI, and a third step from October 2024 has been added.
Added a source to the Kurzweil claim — The entry had a real prediction but no link. It now points to his own restatement of the 1999 forecast in The Singularity Is Nearer, with his wording of what would count.
Corrected Yann LeCun's affiliation — His entry listed him only as an AI researcher. He left Meta in November 2025 and founded Advanced Machine Intelligence Labs in Paris the following month, where he is executive chairman.
Added a published settlement criterion to every claim — Verdicts were previously reasoned at the point of settling, after the outcome was known. Each claim now carries the test it will be judged against, written when it goes on the record and locked from then on.
Restored the original hedged wording on eight claims — Several entries had been shortened in a way that turned "may", "probably" and "I think" into flat assertions. That makes a speaker look more certain, and more wrong, than they were. The wording now matches what was actually said.