Does AI already have human-level intelligence? The evidence is clear

This paper argues that AGIs are already here. I didn't find the central argument fully convincing, but honestly that's part of why I liked it: the diversity of perspectives it opens up raises more questions and provokes more curiosity. It gives us a lot of new perspectives to think about what intelligence even is: human intelligence, artificial intelligence, "general" intelligence, AGI, the elusive thing we might call human-general-intelligence.

The paper starts by revisiting Turing's 1950 imitation game and uses the examples of Olympiad medals, a 2025 Turing-test pass, theorem-proving, and others to claim that the answer to Turing's question (whether machines can display the kind of flexible, general cognitive competence characteristic of human thought, well enough to pass as human to unaware humans) is now "yes." The authors then raise the question: if the answer is yes, why do most expert researchers still hesitate to call this AGI? They argue the reason is partly emotional, partly commercial, and partly conceptual, and in this paper they mainly explore the conceptual aspect. I'd be very interested in learning more about the emotional and commercial aspects. I mean, it makes sense to think about those two. But it'd be very interesting to see case studies, especially on the commercial side.

So the authors do their main work on the definition. First they dismantle any definition strict enough to exclude actual humans, naming four features that are not required: perfection, universality, human similarity, and superintelligence. Then they give a loose definition, "sufficient breadth and depth, anchored by paradigm cases," and defend the vagueness as "a feature, not a bug." With this new definition, they set up a "cascade of evidence" with three tiers (Turing-test level, expert level, and superhuman level) and argue current LLMs already clear the first two. Then the key move: inference to the best explanation. We attribute intelligence to other people just from behavior, so by the same reasoning we should attribute it to LLMs. The final section answers ten objections: they're just parrots, they lack world models, they understand only words, they don't have bodies, they lack agency, they don't have a sense of self, they are inefficient learners, they hallucinate, they lack economic benefits, and their intelligence is "alien." Each gets some version of "that standard would exclude humans too." It closes by putting itself alongside Copernicus and Darwin, another displacement of human specialness.

I noticed that the authors keep using the same move to dismiss each of the four features they say aren't required (perfection, universality, human similarity, superintelligence): humans don't have X either, so we shouldn't require X of AGI. No human is perfect, no human is universal, no human is superintelligent, so none of these can be necessary. It's a clean argument, and it works as far as it goes. But since I wasn't able to trace the source of these four criteria, I have a few questions.

Here's what troubles me. The reason humans fall short of these things usually has to do with our constraints: limited brains, limited time, a single lifetime that only lets you specialize in one or two areas, and so on. And those constraints don't transfer to AGI. So when the authors wave away a feature by saying "humans can't do it either," I think we should ask: can't humans do it for the same reason an AGI can't? Often the answer is no, and that makes the dismissal feel unfair. The bar gets lowered to human level, but AGI isn't built under human limits.

Here are some thoughts on perfection, universality, and superintelligence. I might come back to the human-body part later; at first look I don't have many problems with it, but I've just started reading about why ontic grounding matters, so I may return to it soon.

On perfection. The authors claim perfection isn't required: we don't expect a physicist to match Einstein or a biologist to match Darwin, and few humans have perfect depth even in their own specialty, so perfection can't be required of AGI either. But it's important to note that an AGI is trained on essentially all of human knowledge. Einstein's "plate" had more physics on it and Darwin's had more biology, but an AGI is trained on a combination of both plates, so it should hold the full physics and the full biology, the full extent of every specialty up to our current maximum. Granted, there may be areas the totality of our knowledge hasn't yet covered, and the AGI can't be perfect at that unknown frontier. But within known specialties, it should be. The reason we forgive humans for not being perfect (one brain, one lifetime, you can only go so deep) just doesn't apply here. And notice that what I'm asking for here is really both at once: full coverage across specialties (breadth) and full mastery within each (depth). The authors dismiss these separately, perfection on one side and universality on the other, but for an AGI they may come back together, since the human limits that force a trade-off between going wide and going deep don't bind it the same way.

On universality. The authors argue that no single human can do every cognitive task, and that other species even beat us at some (an octopus controls its arms independently, some insects see light we can't), so general intelligence doesn't require "perfect breadth." But AGIs are standardized in a way humans aren't. Once one AGI is produced, 10 people can use it, 1,000 people can use it; millions of OpenAI or Anthropic customers are all using the same underlying product. Different AGIs might have different functions, but the variation among AGIs is surely far smaller than the variation among humans. So the human-style spread the authors lean on, where everyone excels and falls short in different places, doesn't obviously carry over to AGI. (One side note, not a real objection: this standardization introduces a dimension that barely shows up for humans, namely consistency, and the authors' breadth-and-depth definition has no slot for it. I don't think this counts against their conclusion. If anything, LLMs score well on consistency. I mention it only because it suggests their definition was cut to fit humans, which is the same worry running through all my points here.)

On human similarity. Setting the body aside for now.

On superintelligence. It's interesting how people, by default, seem to expect superintelligence the moment AGI comes up, as if "it's a machine, so it must be extra smart." Where does that expectation come from? My guess is it isn't pure bias: AGI really is superhuman on some axes (speed, memory, never tiring, serving millions at once), and people then unfairly expect it to be superhuman on all axes. If that's right, then "no human is superintelligent, so we shouldn't require it" feels a little unfair, and too quick. The real puzzle isn't whether to demand across-the-board superintelligence; it's how to classify a thing that crushes us on some axes and falls below us on others (counting letters, hallucinating). Which is exactly the "alien intelligence" the authors land on at the very end, so maybe they half-concede this themselves.

Notes (5/26/2026): I won't go through all ten objections one by one at the moment. Several of them lean on the same "humans don't do this either" move I push back on above, and a few (especially the "parrot" objection) I'd rather come back to after reading more on grounding. I might come back and revisit them, especially the embodiment and ontic grounding discussion. Anyway, gotta go to lunch.