Musk promised AI smarter than any human by 2025. He now says 2027.

In April 2024, Elon Musk said on X that AI would be "smarter than any one human probably around the end of next year," and that "the total amount of sort of sentient compute of AI" would exceed all humans within five years, per AI critic Gary Marcus.

In 2026, Musk posted on X that AI "will be able to do anything digital (that doesn't require shaping atoms) at a superhuman level by the end of next year," and followed with a second X post restating the same capability claim. The two predictions are structurally the same: AI that matches or beats the best individual human on a wide range of tasks, delivered in roughly twelve months. The first deadline has already passed. The second is being reported as new information.

"Smarter than any one human" and "superhuman at anything digital" are different words for the same milestone. The first was supposed to arrive by the end of 2025. The second is now promised by the end of 2027. No public reconciliation of the original date has been published alongside the new one.

The Information ran a piece the same day, treating Musk's 2026 posts as a fresh forecast rather than a restatement of an older claim whose deadline had already passed, according to Marcus's Substack. Marcus has framed this as a recurring issue: rewritten predictions being covered as if the prior version had not existed. The Information piece itself is described but not independently verified here.

Marcus is not a neutral party. He is one of AI's most consistent skeptics, and he has been publicly critical of Musk for years. In late 2024 he sent Musk a list of ten tasks, ranging from ordinary household cognition to feats only exceptional humans can perform, and offered a $1 million bet that none would be solved within a year. He says Musk did not respond either time. As of the post, Marcus reports that the count of tasks AI has actually solved is "approximately zero." That metric is Marcus's own, not an external benchmark, and the broader capability question is a separate story.

The capability reality the post does name is narrower and harder to dispute. Even frontier models still produce confident errors on basic factual questions. None has produced a novel, screenplay, or news article of the kind a serious human writer can produce in a working day. Those are not contestable targets; they are the floor of what the original 2024 prediction was claiming was months away.

Musk's language around the predictions is part of the pattern. Marcus notes the hedges in Musk's recent posts read as confidence markers: "for sure," "I am extremely confident," "it's looking quite likely." That register, absolute rather than probabilistic, is the same one Musk has used on driverless cars for the last decade, where confident timelines have repeatedly moved without much press accounting for the prior version.

A prediction like this, from a less famous CEO, would be met with a footnote asking whether the source had a track record on prior timelines. Musk's predictions are often quoted without that footnote. The asymmetry is the story: the same kind of forecast gets different journalistic treatment depending on who issues it, and the difference is rarely explained to the reader.

The cost of treating rewritten predictions as fresh forecasts is not hypothetical. Labor markets, capital allocation, and infrastructure build-outs are being planned against AI timelines that look like Musk's. When the date moves and the coverage does not, the planners are not told that the planning horizon has changed.

AI is plainly improving. The question is what newsrooms owe readers when a CEO prediction has already missed, the same claim is being issued with a new date, and the trade press is carrying the new date as a forecast rather than as a restatement. Either the original prediction has to be reconciled in the same story, or readers should be told the second post is a new deadline for a milestone that was already supposed to be here.