On the Knowledge Problem, Tacit Knowledge, and Artificial Intelligence is an article we need to study. It begins:
In May 2023 the MIT economist Daron Acemoglu, who went on to share the 2024 Nobel prize in economics, asked a question on Twitter. Friedrich Hayek argued in 1945 that central planning fails because no planner can collect and compute all the information an economy runs on. “What if computational power of central planners improved tremendously?” Acemoglu wrote. “Would Hayek then be happy with central planning?”
About thirty economists sent him the same reply, according to Shruti Rajagopalan, who hosts the Mercatus Center’s Ideas of India podcast: Hayek never made a computational argument. He made an argument about the nature of knowledge itself. Acemoglu reached the same conclusion in print that June. In an essay titled “Hayek vs. AI Socialism”, he quoted Hayek’s warning that the relevant knowledge “by its nature cannot enter into statistics” and concluded that “not even an all-powerful large language model” could deal with it. AI socialism, he wrote, “offers only a superficial critique of Hayek.” That distinction is the whole story, and it explains why a model trained on most of what has ever been written still cannot do what a well-trained human mind does.
Thanks for linking the article. It is a good segue into many related works from various disciplines that are all relevant to AI triggered issues we discuss here. Philosophy, economics, political science, governance and regulations — impressive for a mere sales pitch piece (the author makes it at the end). His mentioning of important points of Ryle (know what/how distinction), Polyanyi (tacit dimension) and Hayek (wrongheadedness of central planning/command) are worth it on their own. IMHO important combinatorial considerations are missing — the sheer amount of AI outputs generates a difficult validation problem which is really tough in any domain where the stakes are high (high blast radius of errors as we say in software engineering) like medicine, nuclear and biological research etc. Today one AI is used to validate another A’s solution, but it is very risky.
Dmitri,
I have yet to study the article with care, but it did send me back to Polanyi who I first discovered way back in ’72. His PERSONAL KNOWLEDGE: Towards a Post-Critical Philosophy is rich, fascinating, and surprisingly relevant to out current concerns about AI.
https://polanyisociety.org/michael-polanyi/
Hi Bill
I completely agree with you on Polyanyi. I read his Personal Knowledge more than 30 years ago and have it on my shelves. He is one of those rare 20th century thinkers who are going to stay relevant for a long time.
I am familiar with some of Polanyi’s work, but have read and long admired Hayek. As such, I am pleased to see Hayek’s work showing up again among Bill’s interests. For those few among this blog’s quite astute readers who may not be familiar with him, here is a quick overview. Before he narrowed his focus, Hayek gained wide-ranging exposure to the philosophy of his day amid the intellectual ferment of Vienna (of which the Vienna Circle was a prominent part), and was an early reader of his distant cousin Wittgenstein’s Tractatus (which he later wrote made a “great impression” on him). Over the decades that followed he produced groundbreaking work in economics, the philosophy of science, and the social sciences. He remains widely cited today across such disparate fields as economics, law and jurisprudence, complexity theory, sociology, psychology, and neuroscience, and is easily one of the foremost cross-disciplinary thinkers of the twentieth century. So it is not surprising that something he wrote in 1945 remains relevant to AI issues in the twenty-first.
NB, I am more familiar with Hayek than Polanyi, so my comments will center around Hayek’s essay.
I am not persuaded that AI intelligence is limited regarding the types of knowledge that Hayek identifies. Hayek certainly makes a strong case that neither AI nor a human expert can successfully manage an economy. That does not, however, mean experiential local and personal knowledge cannot be explained and codified for AI use as knowledge-grist in an AI production. In fact, that is precisely what Hayek has done. He has clearly articulated what local and personal knowledge is. The addition of actual local and personal facts would be all that’s necessary for AI to begin its work.
I am more intrigued by the vivid example Hayek offers of an economic agent who must make successful decisions based on partial and incomplete knowledge. That is abduction, an inference/judgment to the best explanation that is underdetermined by the phenomena to be explained. It is a ubiquitous human talent, important in every domain of knowledge and in the course of everyday life. And C. S. Peirce noticed the remarkable thing about it: correct abductions occur significantly more often than random guessing. Given the near infinite hypotheses or theories that might be applicable to a set of facts, this should not be the case. But it is, and those regarded as the most intelligent among us seem to be particularly good at it.
In a post on this blog a little over a year ago – https://maverickphilosopher.blog/index.php/2025/07/14/cognition-hallucination-intelligence-and-ai-1/ – the question was raised whether AI could be trained in abduction. At the time it seemed to me doubtful. Abduction is an optimization strategy involving feel, spontaneity, and intuition; there is simply no rule that could tell an AI how to choose among the multiple theories that might fit the facts the way an intelligent human does. The examples of an experienced accountant who can “smell a rat” in an audit, or a doctor with a stellar record in correct diagnoses, were advanced as things that AI could never do. But at that point, Bill cautioned that “it is risky to assert confidently what computers can and cannot do.”
And sure enough, here we are a little more than a year later and apparently AI is becoming quite adept at audits, medical diagnoses, and abduction in general. I would submit that to the extent AI forms correct abductions on par with or better than human beings, then it is certainly arguable that it is intelligent in the human sense.
Even so, is AI abduction only algorithmic mimicry of human intelligence? And does that make a difference?
This question was posed in last year’s post: whether AI counts as intelligent if it arrives at correct conclusions by a process of thought different from the way humans think. In fact, AI’s process of abduction is quite different. It is computational, based on statistical regularities and patterns, rather than the human, embodied process that is more spontaneous and intuitive. However competent, then, that AI might become, its actual processes of abduction will always remain at best a simulated approximation of human abduction.
And I think that somewhere in the human process of abduction, sentience plays a critical part in the depth of human understanding, meaning, purpose, and value. And It is not established at all that AI is sentient. Here is Grok describing limitations in its abduction processes that seem to implicate a role for sentience in human intelligence:
AI has no “biological instinct or embodiment … [and lacks] genuine curiosity-driven iteration in the wild. Humans continuously test abductions against ongoing experience in the physical and social world. I can simulate that process within a conversation or against known data, but I do not independently act and observe in the same open-ended way.”
This needs a fuller treatment, but for now, I think AI intelligence is real, well established, and immensely useful in many domains, but it is not true human, embodied, sentient intelligence. And the difference between them impacts what AI can and cannot do in critically important ways.