There's a popular argument in defense of how AI is being built. Toddlers also confabulate. They state plausible-sounding falsehoods with full confidence. They don't seem to know what they don't know. Adults confabulate less, and the difference is mostly cultural — habits of mind absorbed from our environment. Therefore, the fact that LLMs hallucinate isn't evidence we're approaching AI in a fundamentally flawed way. They'll grow out of it the way we did.
It's a clean argument and a popular one in the field. It's also wrong, in a specific and instructive way.
The induction that isn't
The argument has the shape of induction but isn't induction. Induction in any rigorous sense — what John Stuart Mill formalized as the Methods of Agreement and Difference in A System of Logic (1843) — requires multiple cases of the same kind, with shared or varying factors that can be identified as causally relevant. What's actually on the table is one biological system developing along a particular trajectory and one artificial system producing superficially similar outputs. The conclusion drawn is that the artificial system will follow the biological one's path. That's not generalization from cases. It's analogy dressed up as prediction.
The move only works if you've already shown the underlying mechanism is shared — that whatever drives toddlers from confabulation to calibrated uncertainty is also operating inside the LLM. The argument doesn't show that. It doesn't gesture at what such a demonstration would look like. The shared output (confident wrongness) is treated as if it were evidence about the generating process. It isn't. Philosophers who study analogical reasoning carefully — Mary Hesse in Models and Analogies in Science (1963), Paul Bartha more recently in By Parallel Reasoning (2010) — have made the same point repeatedly: a sound analogical argument requires that source and target share the relevant causal or structural features that produce the property in question. Sharing the output isn't enough. The inference is only as strong as the demonstrated mechanism connecting the cases. Here there is no demonstrated mechanism.
You can run this move on anything. Calculators do arithmetic — therefore calculators will develop number theory intuition, the way mathematicians did. Parrots produce speech-like sound — therefore parrots will eventually write essays. Animals are smelly and so are humans — therefore animals will eventually develop civilizations. The structure is the same in each case: shared surface property, asserted shared destiny.
Surface behavior tells you almost nothing about whether two systems are doing the same underlying thing or going to the same place.
This is a version of a critique with a long history in philosophy of mind. Hilary Putnam and Ned Block, among others, spent decades arguing that behavioral equivalence doesn't establish cognitive equivalence. Block's "Blockhead" thought experiment — a giant lookup table that passes any conversational test by retrieving precomputed responses — makes the point cleanly: identical behavior, no underlying mind. The toddler argument runs the same fallacy in reverse: identical behavior treated as evidence for convergent mechanism, with no evidence for the mechanism offered.
What growing up actually does
What's happening with toddlers, as best we understand it, is a complicated mix: embodied sensorimotor feedback, years of social correction, evolved priors humans bring to learning, and slow biological maturation of brain regions involved in metacognition and theory of mind. Children develop the ability to distinguish pretend from real, to track the difference between what they know and what they're guessing, to recognize others might know things they don't. None of this is being demonstrated in current LLM architectures. The hallucination problem persists across model generations, across training methodologies, across scale — and it persists despite training on orders of magnitude more cultural data than any human will ever encounter. If "absorbing cultural anti-confabulation habits" were the mechanism, the system with vastly more cultural data should already be the calibrated one. It isn't.
So even inside the argument's own frame, the evidence cuts the wrong way. The observation it wants to leverage — that humans get better as they mature — undercuts the claim, because humans get better through processes that haven't been shown to operate in LLMs.
There's a deeper issue. Even granting the analogy entirely — even granting toddlers and LLMs are doing the same kind of thing — the developmental arc the argument invokes doesn't go where it says it does. Toddlers don't grow out of confabulation into perfect calibration. They grow into appropriately calibrated uncertainty. The mature endpoint isn't a brain that never confabulates; it's one that knows when it's confabulating, hedges, and treats hedging as a feature. Adults say "I think" and "probably" and "I'm not sure." They leave room to be wrong. They maintain a soft layer between knowing and not-knowing because that softness does real work — it's how cognition stays calibrated, how communication preserves the ability to revise, how disagreement routes through "let me check" instead of through certainty.
The wrong target
The toddler-to-adult arc is actually evidence for a different conclusion: cognition matures by acquiring the right relationship to imprecision, not by eliminating it. What gets called "anti-confabulation algorithms" is more accurately a calibrated relationship to one's own uncertainty. That's a different target than "confabulate less." And it's a target benchmarks and accuracy metrics can't capture, because they measure whether outputs are right, not whether the system has an honest relationship to its own confidence.
Which lands the argument where most of these end up. The dominant approach to AI assumes human cognition is the floor, and the goal is to exceed it. But human cognition is the ceiling — the thing being modeled, approximated, possibly approached. When the target is human cognition itself, "perfection" stops being a coherent concept. The goal isn't "be more right than humans." It's "be calibrated like humans — including about your own imprecision."
The toddler argument articulates a dominant view cleanly, which is what makes it worth pushing on. The view says: the gap is one of degree, time will close it, the same processes that make toddlers into adults will make LLMs into reliable systems. We have no evidence the gap is one of degree. The analogies invoked to suggest it don't establish what they claim. And even if we accepted the analogies, the conclusion they point toward isn't "AI will become reliable" — it's "AI will need to become honestly uncertain." Which is a fundamentally different problem than the one most labs are optimizing for.
A8C Ventures is a venture studio building technology for industries where information asymmetry costs people the most.
