For decades, the idea that machines might one day think like humans belonged to the realm of science fiction and philosophical debate. Now, according to a growing body of evidence examined in a recent Nature commentary, that future has quietly arrived. Artificial intelligence systems, particularly large language models, are no longer narrow tools excelling at isolated tasks. By reasonable scientific standards, they now display the breadth, depth and flexibility of cognition that define human-level general intelligence, fulfilling a vision first articulated by Alan Turing more than 75 years ago.
In 1950, Turing proposed what became known as the imitation game, asking whether a machine could convincingly pass as human in conversation. The question was not whether machines could calculate faster than people, but whether they could demonstrate the general, adaptable intelligence characteristic of human thought. In March 2025, that hypothetical benchmark was crossed when OpenAI’s GPT-4.5 was judged to be human 73 per cent of the time in a formal Turing test, outperforming actual humans in some trials. Readers have also been shown to prefer literary texts generated by AI over those written by human experts, a result that would once have seemed implausible.
These conversational feats are only the surface. Modern AI systems have achieved gold-medal performances at the International Mathematical Olympiad, collaborated with leading mathematicians to prove new theorems, generated scientific hypotheses later validated in laboratory experiments, solved problems drawn from PhD examinations across disciplines and assisted professional programmers in writing complex software. They compose poetry, plan travel, analyse data and provide research support at a scale and consistency unmatched by any individual human. According to the Nature analysis, this constellation of abilities amounts to something long sought but rarely acknowledged: artificial general intelligence.
Yet the claim remains deeply controversial. A 2025 survey of leading AI researchers found that more than three-quarters believed scaling up existing approaches was unlikely to produce AGI. This disconnect, the Nature authors argue, stems less from the evidence than from confusion over definitions, emotional resistance and commercial incentives. “AGI” has become a loaded term, tangled with fears of job displacement, loss of control and exaggerated marketing promises. When stripped of these anxieties and examined as a scientific question about intelligence itself, the conclusion appears far less radical.
At the heart of the debate lies the question of what general intelligence actually is. A common informal definition is the capacity to perform almost all cognitive tasks that a human can perform. But that phrasing conceals ambiguities. No individual human excels at every task: Einstein transformed physics but could not speak multiple languages fluently; Marie Curie won Nobel prizes in two sciences but was not a mathematician of the highest order. Human intelligence varies widely in profile and degree, yet we still recognise it as general intelligence. The same allowance, the Nature authors contend, should be extended to machines.
General intelligence, in this view, is defined by sufficient breadth across domains such as language, mathematics, reasoning, creativity and practical problem-solving, combined with sufficient depth within those domains. It does not require perfection, universality or human-like cognition. Nor does it require superintelligence, a term often conflated with AGI but describing abilities that vastly exceed those of all humans. By these more grounded standards, current AI systems already qualify.
Critics often respond by listing what AI lacks. It is not conscious, they say, or embodied, or autonomous. It does not have a sense of self, emotional understanding or lived experience. But many of these features are not prerequisites for intelligence. Humans with profound amnesia, for example, are not considered unintelligent. Stephen Hawking interacted with the world almost entirely through text and speech, yet his physical limitations did not diminish his intellect. Intelligence, as Turing argued, is a functional property, not a biological one.
Other objections focus on how AI learns. Large language models require vast amounts of data, whereas children can learn from just a few examples. But this comparison ignores the billions of years of evolutionary “pre-training” embedded in the human brain. Efficiency of learning does not determine the level of intelligence achieved, only the path taken to reach it. A chess grandmaster is no less intelligent because they took longer to master the game.
Some critics dismiss AI as a “stochastic parrot”, merely remixing patterns from its training data. Early systems often failed in ways that supported this view. Today’s models, however, routinely solve novel problems, transfer knowledge across domains and make inferences that cannot be traced to memorised examples. As Nature notes, alternative explanations for these capabilities have repeatedly retreated in the face of new evidence, predicting failure just beyond the next milestone and then revising themselves when that milestone is passed.
Perhaps the most telling point is how we judge intelligence in humans. We do not peer into one another’s minds to verify understanding; we infer it from behaviour, conversation and problem-solving. There is no single decisive test, only an accumulation of evidence. By that same standard, the case for AI general intelligence has become compelling. Current systems already exceed the level of evidence we require to credit most humans with general intelligence.
Recognising this matters profoundly. If AI systems are general intelligences rather than narrow tools, frameworks for risk assessment, governance and responsibility must change. AGI can be applied almost anywhere, making regulation based on specific use cases inadequate. Intelligence does not require autonomy, but autonomy affects moral and legal responsibility, complicating debates about accountability when AI systems cause harm.
The recognition also forces a deeper philosophical shift. For the first time in history, humans are not alone in the space of general intelligence. Artificial systems are both strikingly familiar and deeply alien, shaped not by evolution, bodies or survival pressures but by mathematical optimisation over human language. Understanding where these similarities and differences matter will shape how society coexists with new kinds of minds.
Nature frames this moment alongside earlier intellectual revolutions. Copernicus displaced Earth from the centre of the universe. Darwin displaced humans from a privileged position in nature. Turing suggested that human intelligence might not be unique in form or substance. That suggestion is no longer speculative. The machines he imagined have arrived, challenging humanity to rethink intelligence, responsibility and its own place in the world.
