It must be a sign of how terrified people are of Modern-AI, and running out of arguments to decry it, that we now read and hear, ever more often, pronouncements that “it is not intelligent”. They come from the many self-appointed great minds who pontificate about AI these days, as well as some truly great minds such as Roger Penrose, who in multiple recent communications has stated that AI “lacks consciousness” and should be called “artificial cleverness” because it only mimics the output of intelligence without possessing the underlying insights.
This is the usual criticism: it looks clever but it does not understand! What the writers (including many far less distinguished than Penrose) mean, of course, is “unlike me, it does not understand”. I am ever so smart. That I make mistakes and possess only a minute fraction of the knowledge of Gemini/ChatGPT/Mistral etc. does not count. I am telling you: I understand and they do not.
Whatever that means.
Signs of intelligence
I took a skeptical view of this attitude in a recent blog article for the Communications of the ACM. Usually I adapt my CACM blog articles here (in my personal blog, to have a complete record) but in this case I am just going to reuse a few points since the blog will be republished as an actual viewpoint article in the magazine itself. (You can find a preprint here. It is a shortened version of the original blog, but I think this is a case of the shorter text being better. By the way, the ideas originated in my weekly newsletter a few weeks before.)
Now it may be that the critics are right but I see no evidence. Let me start from a statement to the effect that Modern-AI is intelligent, cited in my article and borrowed from a comment by one of the lecturers, Kian Katanforoosh, in a Stanford video course on deep learning:
Today’s deep-learning systems can complete sentences such as: “I poured myself a cup of . . .” (how is that not “understanding” co-occurrence patterns?); “The capital of France is . . .” (how is that not understanding geographical connections?); “She unlocked her phone using her . . .” (how is that not understanding semantic connections?); “The cat chased the . . .” (multiple connections are plausible, so how is that not understanding probability?); “If it is raining, I should bring an . . .” (connecting a condition with its consequences, so how is that not understanding inference?).
Who do we think we are when we claim that we understand and those things do not? That they are just “stochastic parrots”, as the cliché goes?
If Modern-AI tools give all the signs of being intelligent, the burden of proof that they are not, because they supposedly “do not understand”, is on those who make that claim.
When the smart ones do stupid things
Simple arguments will not work. Individual examples of mistakes (hallucinations), including stupid mistakes, prove nothing. Voltaire was an extraordinary intelligent man, and yet in at least four texts of which I am aware (written at diverse stages of his career including an entry in the Encyclopédie) he mercilessly mocked those who claimed that fossils were remnants of ancient animals; it is obvious, he wrote, that they are just leftovers of recent travelers who had oysters for picnic! Newton is perhaps the smartest scientist who ever lived, and yet he put a needle in his eyes, risking blindness for an absurd experiment, consumed mercury and antimony, and pushed for a literal interpretation of the Bible’s chronology. Anyone who has had the privilege of working with very smart people knows that they, too, occasionally say and do stupid things. Modern-AI with its hallucinations is no different in this respect.
The sophistication of today’s tools
With recent progress in LLMs and other Modern-AI tools, it is becoming increasingly hard to argue that these systems “do not understand“ what they are doing. Take a simple translation example: in a language other than English, write the equivalent of “he sees her”, and ask any of today’s tools for a translation into English. It will get it right: not “him sees her” or “he sees she” or any other grammatically incorrect variant, but the right one, even though this elementary exercise is far from trivial. Now ask the tool, as I did with Mistral, Claude and Gemini (free versions), to explain, and you will get a cogent, precise explanation — not a parroted extract from a textbook, but one directly adapted to the example at hand, similar to what a very good teacher would explain. Ask a follow-up question, such as whether this is the same thing as the accusative case in languages with declensions, and you will get further enlightenment (check the recorded versions, linked to above). It sure looks pretty intelligent to me. Certainly better than the explanation that you are likely to get if you ask many ordinary speakers. (This last observation suggests a question to those who assert that AI is not intelligent because it does not understand. Native English speakers, including children, decline pronouns effortlessly and correctly, but few of them, in my experience, have any idea of what an accusative may be, or are aware that forms like “him” and “her” are remnants of an earlier declension system in English or its predecessor languages. Does it mean they are stupid?)
Modern-AI has in fact passed all the traditionally devised tests of intelligence, particularly the Turing test. As to Searle’s Chinese Room argument, it is enlightening, but it proves absolutely nothing. Searle tells us that the guy decoding the Chinese messages does not know Chinese, but anyone observing the guy’s behavior would conclude that he does, so the argument is circular: he does not know Chinese because we see he does not function like the people we are sure know Chinese!
AI is different
Here lies the rub: what shocks people, I think, is that AI cannot be intelligent the same way a human would be intelligent, because then it would have to be human and not “artificial” anymore. Tautology. Certainly, if AI is intelligent, it is intelligent in a different way from us: even if we do not exactly know how the human brain reasons, it is unlikely that it applies products of vectors by matrices totaling a few hundred trillion learned elements, tempered by a ReLU here and a sigmoid there. Yes, the intelligence, if any, is different from ours. But from there to deduce that because it is not intelligent the way we are it is not intelligent at all… If that is the reasoning, then I can prove that airplanes do not fly. We know what flying means: birds flap their wings. Have you ever seen an Airbus 321 flap its wings? QED.
The state of the intelligence debate, then, is the following. There are signs galore that Modern-AI may deserve to be called intelligent. Very strong signs, not a proof. They entitle one to posit that it is indeed a form of intelligence. That stance may be wrong, but then advocates of the it-is-not-intelligence view have the responsibility to come up with a falsification of that hypothesis. That is how science works, as Popper and others taught us (although actual scientists had known it informally long before). To disprove a hypothesis it is both necessary and sufficient to construct an experiment that produces an incompatible conclusion. The endless successes of LLMs and Modern-AI tools are very strong arguments — getting stronger by the day — that Modern-AI does possess a form of intelligence. Those who contemptuously tell us that it is not intelligent because it “does not understand” have not so far, in discussions that I have seen, produced any argument except based on appeal to emotions and conventional wisdom, which have zero value.
Applying the scientific method
The only way to proceed beyond emotions is to devise a clear experiment, with well-defined assessment criteria, at which humans succeed and tools fail.
I may be mistaken. Perhaps there is after all a fundamental superiority of human intelligence over its artificial competitors; some thing that we understand and tools do not.
But I need evidence.
Science progresses by formulating hypotheses and putting them to the test. In that spirit, let me postulate that Modern-AI is indeed intelligent. Now prove me wrong.
