Whether AI will surpass human intelligence cannot be answered until "surpass" is defined, and researchers have defined it in at least four incompatible ways: fooling a judge in conversation, scoring higher on a formal measure of ability, exceeding the best people in nearly every field, or outperforming people at most economically valuable work. Under none of these has a machine been shown to meet the standard in full. The 2026 International AI Safety Report, written by more than 100 experts, says that methods for estimating when new AI capabilities will appear remain unreliable.
Key takeaways
- Alan Turing replaced the question "Can machines think?" with a test of conversation in 1950, so the oldest yardstick measures imitation, not ability across the board.
- Legg and Hutter proposed a mathematical measure of machine intelligence in 2007, and it remains a theoretical proposal.
- Philosopher Nick Bostrom's definition of superintelligence requires an intellect much smarter than the best human brains in practically every field.
- OpenAI's charter defines general AI by economic output: highly autonomous systems that outperform humans at most economically valuable work.
- The 2026 International AI Safety Report finds uneven progress and says experts disagree on whether current rates of improvement will continue.
Four ways to define the finish line
The word "intelligence" has no single agreed meaning, which is why predictions about machines passing people talk past each other. Each definition below sets a different bar, and a system could clear one while missing the others.
| Yardstick | Origin | What would count as surpassing people | Main weakness |
|---|---|---|---|
| Conversation test | Turing, "Computing Machinery and Intelligence," Mind, October 1950 | A judge cannot reliably tell machine from person in an imitation game | Rewards convincing talk, not competence at tasks |
| Formal measure | Legg and Hutter, Minds and Machines, 2007 | A higher score on a general measure of success across many environments | Theoretical; not a practical test |
| Superintelligence | Bostrom, "How Long Before Superintelligence?", written 1997, reprinted 2006 | Much smarter than the best human brains in practically every field | Needs a way to compare across every field |
| Economic output | OpenAI Charter | Highly autonomous systems outperforming humans at most economically valuable work | Depends on how work is counted and on the company's own wording |
Turing's substitute for a hard question
Turing's 1950 paper in the journal Mind opens by considering "Can machines think?" He argued that the terms "machine" and "think" were too vague to settle the matter and proposed an imitation game as a replacement. The game is a measure of whether a machine's replies can pass for a person's in conversation. Passing it would be notable, but the test says nothing about whether the machine can do accounting, plan a surgery, or notice that it is wrong. That gap is why later researchers looked for measures that cover abilities instead of dialogue.
Measuring a machine against a general standard
Shane Legg and Marcus Hutter began their paper on universal intelligence from a gap they described plainly: there is no agreed definition of intelligence, particularly for artificial systems that are unlike humans. They gathered informal definitions from experts, extracted the shared features, and turned them into a general mathematical measure of intelligence that could apply to any machine. Their measure ties success to performance across many possible environments. It is an idea for how a definition could be built, and it has not become a test that anyone administers to products.
What the 2026 International AI Safety Report says
The extended summary of the report, published February 3, 2026 and led by Turing Award winner Yoshua Bengio, offers the most careful recent snapshot. It states that AI agents can now reliably complete some programming tasks that would take a human about half an hour, up from under 10 minutes a year earlier. It also says performance stays uneven, that agents remain unreliable on tasks with many steps or unusual features, and that current methods do not give the reliability that many critical settings require.
On the future, the report is cautious. It says that "methods for estimating how and when new capabilities emerge remain unreliable" and that experts disagree on whether developers can keep building more powerful systems at the current efficiency. It adds that if one measured trend continued, with the length of software tasks doubling roughly every seven months, systems by 2030 could reliably complete well-specified engineering tasks taking people several days. Whether that trend holds, and whether it carries over to other fields, is described as unclear.
Why forecasts diverge
- They use different definitions, so one forecaster's "surpass" is another's "match."
- They extrapolate from different measurements, such as test scores, task length, or revenue.
- They differ on whether more computing power and data will keep producing gains.
- They disagree on whether skill in software and mathematics predicts skill in physical or social settings.
Treating a forecast as evidence therefore requires asking which of these assumptions it makes. A date without its definition carries little information.
FAQ
Has AI passed the Turing test?
Researchers have run controlled versions of the imitation game with mixed results, and no study settles the matter for all systems. Even a pass would show conversational imitation, which is one narrow yardstick among the four above.
What is the difference between general AI and superintelligence?
General AI usually means performance comparable to people across a wide range of tasks. Superintelligence, in Bostrom's wording, means an intellect much smarter than the best human brains in practically every field. The second bar is higher than the first.
Can anyone say when AI will surpass humans?
Not reliably. The 2026 International AI Safety Report states that methods for estimating when new capabilities appear remain unreliable, and it describes expert disagreement about future progress.
Does a better chatbot mean AI is nearing human intelligence?
Not by itself. Gains on specific tasks are well documented, but the same report describes uneven performance, with failures on some seemingly simple tasks.
Related reading
- Natural and artificial intelligence compared
- How conversational interfaces developed
- Which industries are least exposed to AI
- Whether AI can evolve beyond human capabilities
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- International AI Safety Report 2026, extended summary for policymakers, February 3, 2026
- Turing AM, Computing Machinery and Intelligence, Mind 59(236), 1950
- Legg S, Hutter M, Universal Intelligence: A Definition of Machine Intelligence, Minds and Machines 17(4), 2007
- Bostrom N, How Long Before Superintelligence?, 1997, reprinted in Linguistic and Philosophical Investigations, 2006
- OpenAI Charter, OpenAI, 2018