When Will AI Reach Human Level? A Question That Hides a More Useful One

At the start of his 1950 paper Alan Turing asked, "Can machines think?" He then argued that defining "machine" and "think" by ordinary usage would turn the matter into an opinion poll, and proposed a practical substitute, the "imitation game". Seventy-five years later we fall into the same trap when we ask about "human level" without saying what we mean.

Which "level" do we mean?

In a paper published at ICML in 2024, researchers proposed grading general intelligence along two dimensions, depth of performance and breadth of capability, and also addressed autonomy, to give researchers a shared language for comparing models and assessing risks. The question then stops being "have we arrived?" and becomes "at what depth, in how many domains, and with how much autonomy?" We explained the difference between narrow and general intelligence in the article "Three Myths About Artificial Intelligence".

What do researchers estimate?

In a survey of 2,778 researchers who had published at leading venues, the aggregate gave a 50% chance that unaided machines would outperform humans at every possible task by 2047, and a 10% chance by 2027. The previous survey's estimate was about 13 years further out, so the estimate itself moves fast. Stranger still, the same survey asked about "every task" and gave 2047, then asked about fully automating "every occupation" and gave a 50% chance at 2116 and 10% at 2037. Two questions that look close gave far-apart dates, which confirms that wording shapes the answer as much as science does.

A review published by 80,000 Hours gathers the estimates of other groups: leaders of AI companies spoke in 2025 of a few years; the Metaculus platform recorded in February 2026 a 25% chance before 2029 and 50% before 2033, while the "superforecasters" in a 2022 survey were further out. The review explains the grounds for caution: differing definitions, incentives to exaggerate in some groups, and the possibility that skill at forecasting near events does not carry over to wholly new ones. These are estimates attributed to their authors, not facts, and knowledge of what has not happened rests with God alone.

And what do measurements say today?

The Stanford 2026 AI Index reports that models exceed the human baseline on several tests, while agents lag at operating a computer: accuracy on the OSWorld test reached about 66.3% in 2025, against 72.35% for humans. The picture is uneven, not a straight line towards a single threshold.

The more useful question: what do we delegate?

Systems do not need consciousness to redistribute work and decisions. And capability follows whoever commands the means of training: the same index puts investment in AI in 2025 at about $285.88 billion in the United States, against about $20.92 billion in Europe and about $12.41 billion in China, according to its own counting method. The matter concerns sovereignty and the distribution of benefits as much as technology.

The OECD principles, adopted in 2019 and updated in 2024, count accountability and transparency among the value-based principles. A decision that touches people's lives still needs a human who answers for it. Before delegating a matter to a system, ask:

• What is the harm if it errs, and who bears it?
• Who reviews the decision before it takes effect, and by what standard?
• Can we explain its reasoning and appeal against it?

We should take the means, in learning and in governance, and rely on God for what we do not control. As for "when?", it is fair to follow it with a critical eye, and to ask alongside it, "with what responsibility?"