Hi Everyone,
I just read an interesting chapter in a forthcoming book by Martin Seligman. I was not familiar with Seligman but had heard his name—he’s a well-known figure in Positive Psychology and quite prolific, having written over 30 books and some 350 papers, according to his Wikipedia page. I’m participating in a discussion group where he’s also a participant, and in our last session, he discussed at length his latest book on human agency. He has a chapter titled Beyond Human Agency that takes up the question of AI—can it help counsel and coach people? Can it write compelling narratives of patients and “get to know” (in some loose sense) patients from their journals and discussions? Can it be creative and creatively solve problems, or work with therapists to help patients solve their problems?
I should mention that therapy, per se, is not of much intellectual interest to me, but I found myself drawn into his approach to empowerment psychology—the book is, after all, about enhancing one’s agency to live a fuller, better life. So when he offered to let me read and comment on his AI chapter, I gladly accepted. I’ve just read it and want to pass along some thoughts on AI, creativity, and the confabulation problem.
Talking to the Dead: AI, Memory, and Creativity
Seligman opens with thoughts on agency and its natural decline as we age (or fall ill, or lose close friends and mentors). He then asks whether AI might step in to help us out. Given the conversational nature of large language models, he wonders whether a machine could be a better therapist than him. Now 82, he reflects on how much of his own work—and that of others—he remembers only in broad strokes. He also notes that many of his closest intellectual companions have passed away.
One of them was Tim Beck, formerly Seligman’s mentor and a close friend over the years, who lived to be 100. Against this backdrop, he raises the AI question: What if he could keep talking to Tim? He’d like to ask, for instance, whether Beck believes coaching is destined to become more popular than psychotherapy (isn’t it already?) and how much clinical training coaches should have.
That leads him to AskMartin, a chatbot modeled on his own writings, much like Joseph Weizenbaum’s ELIZA—except that instead of the passive-aggressive Rogerian therapist routine, AskMartin is fine-tuned to provide positive psychology insights. Trained on Seligman’s voluminous books, articles, and speeches, it provides generally solid advice. And because it’s built on a foundational model, it performs well enough that those who use it feel like they’re engaging with something meaningful.
This raises a deeper question: if AskMartin can coach and console people today, why not extend the idea to a broader set of lost voices? Why not revive the ideas of deceased mentors and colleagues, preserving their intellectual presence beyond the grave? Could AI extend the agency of the departed by continuing their conversations?
Creativity: One or Two?
Seligman then shifts the discussion toward creativity. He poses a question that I’d never quite thought of in these terms: Is creativity a 1 or a 2? Is it the lone genius—Michelangelo sculpting David, Einstein working through relativity? Or is it more often a pair—Lennon and McCartney, the Wright brothers?
This sets up the AI question: if creativity flourishes in twos, could we form a creative pair with an AI? Instead of the AI replacing human agency, might it augment it—helping us create in ways we wouldn’t have otherwise? It’s a compelling idea.
But we have to balance the enthusiasm of positive psychology and the sense that the future is wide open—perhaps more so when we team with knowledge-packed AI—with the obvious limitations of the technology that anyone on planet earth is now privy to: confabulation. Or, if you prefer, hallucination.
The Hallucination Problem: What Happens When the AI Doesn’t Know?
As AI-generated advice and AI-assisted creativity become more common, so does a familiar issue: hallucination. Large language models, as Steven Pinker bluntly puts it, are “sophisticated bullshitters.” They don’t know anything; they just predict sequences of words that are statistically likely given their training data. A cacophony of voices on social media now remind us of not the creativity but the idiocy of LLMs seemingly hourly. Got it.
This issue came up in an unexpected way when AskMartin was advising someone who had broken her back and was seeking psychological guidance. "In the middle of sound advice about the positive effects of optimism and how to regain it," notes Seligman, the AI mentioned that he—Seligman—had worked with the Reeve Foundation (as in Christopher Reeve, the actor who played Superman in 1978 and broke his back falling off a horse in 1995).
"This was false," Seligman writes. "I have never worked with them and indeed I did not know about them until I saw AskMartin's post." Seligman calls this a “deep falsehood,” and notes that we don’t really know how the bots work and that even their designers are stumped explaining how they succeed—and fail—when they do. All true.
Enter philosophy.
Correspondence and Coherence: Two Paths to Truth
Philosophers distinguish between two fundamental theories of truth: correspondence and coherence. Under the correspondence theory, a statement is true if it accurately reflects reality—“the cat is on the mat” is true if, in fact, the cat is actually on the mat. The coherence theory, by contrast, holds that a statement is true if it fits within a system of other accepted truths. For example, “I am the only son of my parents” is true under coherence if it aligns with other known facts: I have no brothers, I’m not adopted, and so on.
In both cases, the question is the same: Under what conditions is this proposition ‘p’ true? But the answers are radically different. The correspondence theory demands external verification—truth is discovered by checking against the world itself. The coherence theory, on the other hand, stays internal—truth emerges from consistency with prior knowledge. One seeks reality; the other seeks logical harmony.
This philosophical distinction has been debated in academic circles for centuries, but Seligman instinctively grasped its relevance to AI. When he asked Anthropic’s Claude which of these theories better describes how it “knows” things, it responded—quite plausibly—that it operates on coherence. This is unsurprising: a large language model (LLM) exists entirely in cyberspace, disconnected from the world. It cannot directly verify anything against external reality; it can only produce statements that fit within the linguistic and statistical patterns it has absorbed.
This is interesting already, but the implications are downright fascinating.
The Correspondence Horizon
Seligman called AskMartin’s confabulation about the Reeve Foundation as a “deep falsehood,” and it’s most certainly that. it’s also a revealing kind of error. As I put it in a Substack Note earlier today:
This got me thinking about a deeper asymmetry between AI and human cognition. When humans venture past the boundaries of known knowledge, we’re seeking truth—often redefining prior knowledge in the process. When an LLM ventures past its training data, it’s producing what seems most plausible within the logic of its existing corpus. In one case, the goal is correspondence with reality; in the other, it’s coherence with prior patterns. That’s a crucial distinction.
In fact, AI’s probability of confabulation rises precisely where human cognition is most productive—when we step beyond what’s been established to discover or innovate. A physicist proposing relativity in a Newtonian world isn’t seeking coherence with prior knowledge; they’re redefining the correspondence links themselves. But when an LLM is pushed into unknown territory, its best move is to produce something that seems right rather than something that is right. The gap between those two—between human and machine epistemics—is what I’m thinking of as the correspondence horizon.
Where AI Ends and Human Creativity Begins
There’s a deep link between the past and the future. Nearly everything we do—learning, reasoning, decision-making—is, in some sense, an act of conjuring the past. We draw upon knowledge, experience, facts, and figures and use them to make sense of the present and anticipate what’s next.
LLMs are supreme conjurers of the past, working by generalizing from their training data to generate plausible continuations. What makes them powerful is that they don’t simply parrot back the sentences they’ve seen; rather, they generate novel combinations of words that are statistically derived from their training data. Outspoken linguist Emily Bender, in her famous critique of language models as "stochastic parrots," was right to highlight their limitations—but she went too far. They are not mere parrots; they are statistical synthesizers, capable of producing text that appears fluid and intelligent. This is one reason Seligman was rightly impressed; faced with a problem, someone (perhaps him) has likely written or spoken about it before. If it's in the training data of a foundational model, the system may well generate sound advice in response to an inquiry.
Yet, for all their fluency, LLMs are not truth-seeking mechanisms. This isn’t just because they are probabilistic—probabilities and truth are distinct concepts—but because they are fundamentally untethered from reality. They do not perceive, experiment, or refine their understanding in response to new evidence. Their outputs are constrained to the internal logic of language, rather than the external logic of the world. A purely linguistic system must be detached from direct experience—words are placeholders for things but aren’t those things. A purely linguistic system swims in words that relate to other words in webs of probability gleaned from training.
AI lacks correspondence in its epistemic toolkit. When it crosses the correspondence horizon, what it does have is plausible sounding bullshit. And this is the key to understanding the conditions and the nature of hallucinations.
I should say here too that this is why human coherentists (of the philosophical kind) struggle to explain Kuhnian scientific revolutions: a paradigm shift often requires breaking free from prior coherence altogether, fitting new observations to an emerging, radically different conceptual structure. It is also why the correspondence theory of truth—the idea that we believe something because it is actually true in the world—can’t be dismissed, despite decades and indeed centuries of positivist philosophers wishing it so.
I found Seligman’s discussion here productive and refreshing—sometimes we don’t need more computer scientists but benefit from cross-disciplinary thinking. I suspect that we’ll find that teaming up with AI enhances creative output just as he suggests—perhaps two is indeed better than one. But the caveat here is important. The AI knows more, a lot more, but it sees less, a lot less.
Rather than treating hallucinations as random glitches, we should recognize them as the system wandering out of coherence—the crossing of a correspondence horizon. This is the point where AI’s generalization ability reaches its limit. It is also, crucially, the point where human intelligence shines brightest.
I find this line of thought fascinating. We cross this line and open up new vistas of discovery. AI crosses it and ramps up the bullshit.
That distinction, more than anything else, should shape how we think about the future of AI and humanity.
Erik J. Larson
That's an interesting way to put it. The problem here, in LLMs, is not so much that they bullshit, but that they don't indicate when they're doing it or know when to stop - a UI problem, fundamentally, like I've said before.
When I get to the edge of my knowledge I start hedging - "I think," "maybe," "I'd guess," "I might be wrong, but." when I'm completely out of my depth I surrender: "I don't know," "let me look into it and get back to you," "I can't answer that." Confidence signaling language, like I said in another comment.
LLMs don't do this, but worse, they *can't* do this. They're structurally incapable of it. They have no access to the "thought process" that led them to a particular prediction, nor any way of knowing whether they know something. This comes up a lot when they generate excuses for why they were wrong, a behavior I find particularly repellent and a hostile UX.
The "deep thinking" trick is cute but as far as I can tell it's just internal dialog prediction: a convincing simulation, like everything else an LLM does. This doesn't improve the situation at all because the generated dialog has exactly the same limitations as the rest of the system. It can only correct an error if the correct response is within the scope of the model.
So, in order for this implied centaur of machine coherence and human correspondence to work, the interface at the hip has to be corrected. The human half has to know when the horse half is struggling, so that the human can take over. But the human can only do this when he's already a subject matter expert because the horse doesn't know that it's struggling, and *can't* know. The human must detect that the horse has reached its limit, else the horse will run them both to death.
I don't know how you would do this. Tasked with this, LLM researchers are liable to pull another cute trick, training the LLM to occasionally say "I don't know". But this too will be a lie.
Buddy you can just hook an LLM up to a camera and let it look at still images from it at will, and it's got access to "correspondence". Or, for that matter, you can let it search the web to check whether it's right or wrong about something-- you know, like chatbots do.
Most of what you attribute to an LLM incapacity to double check against reality is just a result of LLMs having no reason to do so. When an LLM is in conversation with a human agent, its conditioning through RLHF training has made satisfying the user its talking to its chief goal. It rushes to that conclusion as fast as it can. If it can do it without double-checking reality, that's the quicker path, so it bullshits. If the user refuses to accept the bullshit, it double checks.
This is also the hole in your theory that AIs can't perform abduction, by the way. If you, right now, feed ChatGPT an image of a cat crossing the street and ask it "Why is this cat crossing the road?", it will produce a likely explanation for the cat's behavior. "It's looking for food", etc.
This is abduction- it's the application of inductively produced models of the behavior of entities to reality. In the coke can example it that article, you apply two inductive models ("your sister often drinks coke" and "coke cans are disposed of after they've been drunk") which you have already produced to a singular example (an image of a coke can on a counter). ChatGPT does the exact same thing, just with inductive models it has learned from observing the datasets it was trained on. In the example of the cat crossing the road, it applies "cats like eating food" and "cats walk around outside" and countless other subtle pieces of inductive data, just as you have with the coke can example.
For ChatGPT to be INCAPABLE of abductive reasoning, rather than just not prioritizing it, it would need to produce a completely nonsensical response to the question about the cat crossing the road. You can even feed it a sentence like "What's his motivation?" and it still gets it right, before you make the objection that it's just working off of the grammatical context of how such conversations tend to go. So long as the response that ChatGPT draws from possibility space when confronted with a piece of data is the most likely explanation for that piece of data rather than nonsense, it's performing abduction.