“By reasonable standards… we have artificial systems that are generally intelligent.”
A recent Nature article makes a direct claim: artificial general intelligence may already have been achieved.
In “Does AI already have human-level intelligence? The evidence is clear”, the authors argue that today’s large language models exhibit human-level intelligence across a wide range of domains.
I went into the article skeptical. What kept me reading was how the argument comes together and how carefully the authors define what they mean by intelligence.
What counts as artificial general intelligence?
The argument begins with a simple question: what do we actually mean by artificial general intelligence?
Over time, the definition has picked up additional expectations. It now often implies perfection, complete coverage across domains, and consistently superhuman performance. That makes it difficult to compare to anything human.
The authors take a more grounded approach. General intelligence is the ability to handle a wide range of cognitive tasks at roughly a human level. That includes language, reasoning, mathematics, and problem solving. With that definition in place, the discussion becomes much more concrete.
“A definition that excludes essentially all humans is not a definition of general intelligence.”
Human-level intelligence across many domains
“No single test is definitive, but evidence accumulates.”
Models can pass exams, write and debug code, assist with research, generate ideas, and move across domains with a flexibility that we had not seen before.
It is easy to push back on individual examples. I found myself doing that while reading. What is harder to dismiss is how wide the capabilities are when you step back and look across them. Taken together, they point to human-level breadth in a way that no single example does.
Why AGI standards have shifted beyond human intelligence
One idea that stood out to me is how much the definition of AGI has changed over time.
The bar has become flawless reasoning, complete coverage across domains, and even major scientific breakthroughs.
We don’t judge human intelligence by those standards.
Humans make mistakes, rely on imperfect reasoning, get things wrong, have limited areas of expertise, and most never produce major scientific or artistic breakthroughs.
If a definition of AGI excludes humans, then the definition itself has moved away from the concept it was meant to capture.
Common objections to calling today’s AI AGI
“There would be great opposition [to thinking machines] from the intellectuals who were afraid of being put out of a job. It is probable though that the intellectuals would be mistaken about this.”
Alan Turing, Intelligent Machinery, A Heretical Theory, lecture to the 51 Society, University of Manchester, c. 1951.
The objections are familiar and worth taking seriously.
The “stochastic parrot” argument suggests that models are simply remixing training data. That explanation starts to break down when models write working code for new problems, respond to unfamiliar prompts, and apply knowledge in contexts that were not part of their training data.
There is also the question of whether these systems actually understand what they are saying, or whether they are only producing convincing language. In practice, these systems can explain concepts, apply them correctly, and adapt them to new situations. That begins to look very close to what we mean by understanding.
Embodiment is another concern. Intelligence is often tied to physical interaction, but we do not require that in every case. We would still recognize intelligence in someone who cannot act physically yet can reason, communicate, and engage with ideas. A mathematician working entirely in symbols, or someone reasoning through complex problems in conversation alone, would still be considered intelligent. That makes it harder to treat physical embodiment as a requirement for intelligence.
Does AGI require autonomy or agency?
This is probably the strongest remaining objection.
These systems do not set their own goals or act independently. They respond to prompts. That difference is real.
“Autonomy matters… but it is not constitutive of intelligence.”
The authors argue that autonomy is separate from intelligence. This is where I start to hesitate. The ability to form goals and act on them feels like a core part of what we mean by intelligence.
Why the AGI debate matters for computer science education
Even if the label remains debated, the shift in capability is already showing up in practice.
I wrote about this in an earlier post on AI and the future of programming, where the focus was on how programming is moving toward system design and domain understanding.
It also affects how we think about education. In programs like the CU Boulder Applied Computer Science Post-Baccalaureate program, there is increasing emphasis on working effectively with AI systems and understanding how they fit into larger workflows.
Has AGI already been achieved?
The authors are making a direct claim that artificial general intelligence has already been achieved. You can disagree with the label, but the argument itself is difficult to dismiss.
These systems are operating across domains with a breadth and flexibility that used to define the problem itself.
At a minimum, this forces a more careful comparison between human and machine intelligence. If we continue to say that AGI has not been achieved, we need to be clear about the standard we are using and whether that standard is one that humans actually meet.
“There would be plenty to do in trying to keep one’s intelligence up to the standard set by the machines, for it seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers.”
Alan Turing, Intelligent Machinery, A Heretical Theory, lecture to the 51 Society, University of Manchester, c. 1951. (emphasis mine)