Scientists claim artificial general intelligence is already here

Leading scientists claim that the fabled “general-purpose AI” is already here – we simply fail to notice because we are judging it by the wrong standards.

While technology giants still grandly announce an imminent breakthrough towards artificial superintelligence, philosophers and computer scientists are taking the argument a step further. In their view, today’s systems, including ChatGPT, have achieved what many have pursued for decades: artificial intelligence at a general, human level. The intriguing question is no longer simply when it will arrive. It is whether we would recognise it if it had existed all along.

What researchers mean by “artificial general intelligence”

Two terms are frequently conflated in this debate: everyday artificial intelligence and so-called “Artificial General Intelligence” (AGI) – AI capable of performing at a human level across many very different fields. For years, AGI was regarded as a distant science-fiction ambition.

A group of researchers spanning philosophy, linguistics, computer science and data science is now openly challenging that assumption. Their central argument is that we define intelligence too narrowly and through an overly human-centred lens. As a result, we overlook the fact that current models already do much of what we have long demanded as evidence of “real” intelligence.

The authors argue: If we accept humans as intelligent beings despite being limited, fallible and specialised, we must apply the same standards to machines.

Rather than looking for a flawless, all-knowing superbrain, they advocate a more grounded definition. Artificial general intelligence exists when a system demonstrates abilities within the range of human experts across a broad set of tasks – no more and no less.

AGI is not superintelligence

A key part of the discussion is the distinction between two levels:

• Artificial general intelligence (AGI): systems able to work at expert level in numerous fields, comparable with skilled people.
• Superintelligence: systems that substantially outperform humans in virtually every cognitive domain.

According to the researchers, current large language models (LLMs) already provide many examples of expert-level performance, from programming and law to specialist medical questions. That, they argue, is precisely the benchmark for AGI. Spectacular, superhuman intelligence would be an entirely different stage, and one that may still lie ahead.

Has the Turing test already been passed?

The Turing test is a historic benchmark for “real” AI. Alan Turing proposed it in 1950: if, in a written conversation, a person can no longer reliably tell whether they are communicating with a human or a machine, the machine counts as intelligent.

Modern chatbots pass this test in many situations. In blind tests, users often classify ChatGPT and comparable systems as “human” more readily than actual people taking part in the test. Only a few years ago, such a result would have been treated as clear proof of strong AI.

By the traditional standard of the Turing test, current systems would already be recognised as fully fledged, intelligent conversational partners – yet the bar is being raised retrospectively.

This is exactly where the researchers’ criticism begins: whenever AI satisfies an old criterion, the public shifts the goalposts. Intelligence thus becomes a promise that continually recedes from view.

Common objections to AGI – and why they are less secure than they seem

In their analysis, the scientists address a wide range of familiar criticisms directed at current AI models. Many are well known from public debate:

“They are only stochastic parrots”

One common claim is that language models merely repeat patterns from their training data without genuine understanding. The researchers counter that these models are increasingly solving tasks that did not occur in their training, including entirely new mathematical problems and difficult logic puzzles.

They demonstrate transfer learning: knowledge from one domain helps them tackle tasks in another. This flexible linking of content has long been considered a core characteristic of intelligence.

“There is no real intelligence without a body”

Another objection is that humans have bodies, experience the world through their senses and act within it. Software alone, the argument goes, can therefore never truly “understand” what words mean.

The authors disagree here too. They point to advances in multimodal models, which process not only text but also images, audio and video. Such systems can estimate physical consequences, plan movement sequences or draw logical inferences from visual scenes.

At the same time, increasing numbers of robots are being connected directly to AI models – an area researchers call “Physical AI”. With every advance, the link between digital intelligence and the physical world becomes stronger.

“There is no general intelligence without autonomy and a biography”

It is often argued that a system requires enduring goals, a stable identity and something resembling a life story before it can be considered truly intelligent. The researchers take a less restrictive view: intelligence is shown primarily through behaviour and problem-solving ability, not through whether a machine remembers its “yesterday”.

Anyone who recognises intelligence only when it comes with human consciousness, emotions and a life story creates a definition that excludes machines by design.

The issue of consciousness therefore remains unresolved – but the authors do not consider it essential when classifying a system as AGI.

What about notorious AI hallucinations?

One of the strongest arguments against AGI claims remains hallucination: AI models invent facts, cite sources that never existed or gloss over important details. This still happens frequently today.

The researchers acknowledge the problem, but place it in a different context. They note that humans, too, are prone to faulty reasoning, distorted memories and polished lies. In their view, a high error rate does not automatically mean intelligence is absent – only that it is limited and unreliable.

Current studies nevertheless show that hallucinations may even increase in certain scenarios. According to statements from OpenAI, even future models such as a possible GPT-5 are still expected to contain serious errors in roughly one in ten answers. For critical uses such as medicine, law or infrastructure, this remains a serious risk.

Aspect	Human	Today’s AI
Breadth of knowledge	Highly limited, specialised	Extremely broad, but uneven in depth
Sources of error	Bias, forgetting, emotion	Training data, model limitations, hallucinations
Learning speed	Slow, requiring little data	Fast, requiring enormous amounts of data
Explaining its own decisions	Subjective reasons, often incomplete	Statistics that are difficult to interpret

Why our definition of intelligence may be the real problem

Perhaps the authors’ most provocative claim is this: it is not AI that is lagging behind, but our concept of intelligence. We tie it so tightly to human experience that every machine-based form begins at an unfair disadvantage.

Humans are considered intelligent despite being forgetful, constantly making mistakes and possessing no expertise at all in many areas. At the same time, we demand near-error-free performance from AI systems; otherwise, we deny them even basic intelligence.

A clear anthropocentrism lies behind this: we treat ourselves as the benchmark and ignore that intelligence can take different forms. An email spam filter “feels” nothing, yet it recognises patterns very reliably. A chess program does not understand human emotions, yet it can defeat every grandmaster. And large language models cannot genuinely enjoy a weekend, but they can write fluent prose, program software and solve specialist tasks.

Why technology leaders prefer to talk about superintelligence

It is revealing how major companies frame the debate. Figures such as Mark Zuckerberg increasingly speak of “superintelligence” rather than artificial general intelligence. That automatically moves attention into a distant future, away from the question of whether current systems already represent a new form of intelligence.

This linguistic shift has consequences: if superintelligence becomes the new grand objective, today’s models appear to be harmless intermediate stages. On one hand, this reduces the sense of urgency around risks; on the other, it intensifies the hype surrounding what is supposedly about to arrive.

What artificial general intelligence means for everyday life

Whether or not one agrees with the researchers, their arguments have direct implications for politics, regulation and the workplace. If we view AI systems as generally intelligent actors, we need different rules from those used for simple tools.

• Responsibility: Who is liable when an “intelligent” machine acts largely independently?
• Transparency: How much access to training data and model architecture is required to ensure decisions remain understandable?
• Shift in expertise: Which jobs change when expert knowledge is partly outsourced to machines?
• Education: Must pupils and students learn to work with a kind of “digital co-thinker” instead of merely memorising facts?

In everyday practice, this means that anyone working with AI systems now operates increasingly in a hybrid space. The machine handles some tasks almost entirely on its own, while others require close human oversight. The challenge is to assess realistically what these systems can do – and where their blind spots remain.

Terms and examples that make the debate more tangible

Many of the buzzwords sound abstract, but become clearer when viewed through typical applications:

• Large language models (LLMs): These are based on billions of text fragments and learn how language works statistically. They power chatbots, coding assistants and automated summaries.
• Multimodal AI: This combines text, images, audio and video. A model might analyse a photograph while also producing a written description or answering a question about it.
• Physical AI: Robotic arms, domestic robots and autonomous vehicles use AI to understand their surroundings and take action.

A practical example: a modern system could read the text of a construction guide, identify the correct tool in a photograph, demonstrate in a video how to hold it, and explain safety rules at the same time. To the user, this feels like a combination of expert, teacher and assistant – without a person working in the background.

It is precisely scenarios such as these that support the argument that we may already be dealing with a form of artificial general intelligence, one that simply looks different from how it was imagined in earlier science-fiction films. Whether or not we apply the label “AGI” does not alter the fact that our understanding of intelligence is currently expanding radically – and that we must learn how to live alongside this new kind of participant in society.