People who write predictions get scored. People who say them don't.

Jakob Nielsen published eighteen predictions for 2026 on his Substack in January. Six are now on this record, and all six carry a deadline and a test that will decide them.

Elon Musk spent three hours on the All-In podcast saying things about the future. Seven of those are on the record. One carries a deadline.

Sam Altman, an hour with Alex Kantrowitz: five records, none scoreable. Nate Silver, writing: six for six. Rob Toews in Forbes: five for five. Casey Newton at Platformer: five for five.

The pattern held every time I checked it. Predictions that were written down compiled into a test almost always. Predictions said out loud almost never did.

It isn’t that podcasters are vaguer

That was my first assumption and it is wrong. Musk was specific. “Your car is going to feel like it is sentient by the end of the year” has a date in it. So does “the phones that are able to use the spectrum that was acquired probably start shipping in around two years.” Altman said “we’ll triple our compute again next year,” which is a number and a date.

None of those can be settled. Not because they are vague — because nothing publishes the answer. No series reports whether a car feels sentient. Nobody discloses OpenAI’s compute. The claim is specific and unresolvable at the same time, which is a combination I did not expect to find so often.

Writing seems to fix this by accident. Typing a prediction into a numbered list with a probability beside it makes you reach for something you could later point at — a benchmark, a price, an election, a filing. Saying it into a microphone makes you reach for whatever lands.

Why it matters if you keep score

Half of what gets called a prediction cannot be graded, and the half that can is not spread evenly across people. It clusters in writers. Any track record assembled from interviews is measuring who chose checkable phrasing, not who was right.

This registry holds 101 records. Fifty-one carry a deadline. The other fifty are real, sourced, dated statements that will never be marked correct or incorrect, and pretending otherwise is the easiest way to make a number lie.

The practical version, if you are the one predicting: name the source. Not “inflation will fall” but “CPI below 2% on the December print.” It costs four words, and it is the whole difference between a claim someone can check and a claim that gets remembered however people feel about you later.

The AI predictions and Musk’s record are both public if you want the working.