Is Human-Level AI Really Almost Here?

You cannot read about AI for long this summer without bumping into a prediction that the machines are about to match us at everything. When Demis Hassabis published his framework for governing frontier AI this month, he framed the whole exercise around the belief that artificial general intelligence is probably only a few short years away. He is not alone. Anthropic's Dario Amodei has said publicly that AI systems could be broadly better than humans at almost everything by 2026 or 2027. Hearing it often enough, you would be forgiven for treating imminent AGI as settled fact. It is not.

Start with what has actually changed, because something has. Forecasters really have pulled their estimates forward. The community prediction site Metaculus now puts a 25 percent chance on AGI by 2029 and roughly even odds by 2033, a dramatic compression from a median guess of around 50 years as recently as 2020. Analysts tracking these forecasts describe a genuine shift in expert sentiment, driven by how far large models have come in a short time. That part is real and worth taking seriously.

Where the confidence breaks down

The trouble starts when a shifted probability gets reported as a countdown. Two problems sit underneath the headline numbers.

The first is that nobody agrees on what AGI even means. There is no accepted definition of the threshold and no agreed test for when a system has crossed it. A claim that something will arrive by 2027 is hard to check when the something is undefined. Different forecasters are often predicting different things while using the same three letters.

The second is that the loudest optimists are not neutral. The people promising human-level AI within a couple of years tend to run the companies raising money to build it. That does not make them wrong, but it does mean their timelines carry a commercial charge that a careful reader should weigh. Prominent researchers including Yann LeCun and Gary Marcus argue the opposite case, that today's architectures will not reach general intelligence at all without fundamental new ideas, and that the current scaling path has limits the boosters prefer not to dwell on.

What is fair to say

So treat sweeping arrival dates as marketing-adjacent forecasts, not as reporting. The honest summary is a wide spread of outcomes, with a lot of serious people clustering their guesses somewhere in the 2028 to 2035 window and long tails on both sides. That is a real possibility worth planning for, and also a long way from a promise.

This is the same pattern we keep flagging in this section. Impressive capability gets stretched into claims the evidence does not support, as with the idea that chatbots are becoming conscious or that a bigger context window means perfect memory. Human-level AI may well be coming. Whether it lands in three years or thirty is a question no one can answer yet, and confidence is not the same thing as evidence.