The Smuggled Ideas

This is NOT an "AI and cognitive science are cousins" article. That line has been beaten to death with a respectable academic stick. Everyone nods along, someone mentions connectionism, someone else brings up Marr's levels, and we all go home feeling interdisciplinary.

The real story is weirder. The two fields keep smuggling oddly specific tricks across the border, often from papers that looked provincial or niche at the time. Not grand unified theories. Not keynote-worthy paradigm shifts. Just quiet little heists where one field picks the lock on another's toolkit and walks off with something useful.

Here are four of my favourites — two flowing each way across the border. What makes them interesting isn't the borrowing itself, but that nobody at the time would have predicted any of it.

1. Baby Words → Multimodal Grounding

Smith & Yu (2008): Cross-situational word learning

The original paper was about something beautifully specific: how do 12- and 14-month-old infants learn which word goes with which object when every individual scene is ambiguous? Their answer was statistical accumulation. Babies don't solve the mapping cleanly in a single encounter. They track co-occurrence statistics across many messy, individually uninformative scenes, and the right mappings emerge over time.

At the time, this was a developmental psychology result. It mattered to people who cared about word learning, and those people were not, by and large, building neural networks.

But the logic of the finding was a lovely little bomb under the fantasy that grounding needs neat supervision. If infants can learn word–referent mappings from ambiguous, noisy, unsegmented input — no explicit labels, no pointing teacher, no clean training pairs — then maybe machines don't need pristine supervision either.

That idea migrated. It shows up explicitly in AI work on unsupervised language grounding and in multimodal neural models that learn word–referent mappings from ambiguous paired input. The modern contrastive learning paradigm — think CLIP and its descendants — is not a direct descendant of Smith & Yu, but it breathes the same air. Learning from co-occurrence across noisy multimodal scenes, rather than from curated label–image pairs.

What the receiving field improved: Scale. Smith & Yu's infants saw a handful of objects across a handful of trials. Modern multimodal models do the same trick across billions of image–text pairs. What got preserved was the core insight that ambiguity isn't a bug to be eliminated — it's a training signal to be aggregated.

What got lost in transit: The developmental trajectory. Infants don't just accumulate statistics — they bring changing attentional biases, social cues, and embodied experience to each scene. The AI version kept the statistics and dropped the developmental scaffolding. Whether that matters depends on what you think grounding actually requires.

2. Optic Flow → Drone Landings

Lee (1976): Time-to-collision and tau

This one is gloriously unlikely. David Lee's 1976 paper was ecological psychology at its most elegant: he formalised how an agent could use the rate of expansion of an object's image on the retina — a quantity he called tau (τ) — to estimate time-to-collision without computing distance or speed explicitly. Tau gives you a direct perceptual variable for "when will I hit that thing?" and Lee showed it could explain how people brake, how birds land, and how gannets time their wing-folding dives.

It was, in other words, a perception–action paper about braking and optic flow. The kind of thing you'd encounter in a sensory ecology course, nod approvingly at, and never think about again unless you were interested in gannets.

Then the roboticists found it.

Tau-based guidance laws now appear in work on UAV descent, autonomous landing, docking manoeuvres, and general robot motion control. The appeal is obvious once you see it: tau gives you a control law that doesn't require an explicit world model. No range-finding, no depth estimation, no 3D reconstruction. Just the rate of visual expansion, and a simple coupling to motor output. The same trick evolution discovered for gannets turns out to work rather well for quadcopters.

What the receiving field improved: Robustness and generality. Lee's tau was a theoretical analysis with some behavioural evidence. Roboticists turned it into implementable control laws, tested it in simulation and real hardware, and extended it to situations Lee hadn't considered — like spacecraft docking, where the optic flow geometry gets exotic.

What got lost in transit: The ecological context. Lee's point wasn't just "tau is useful" — it was that organisms can perceive affordances directly, without building internal representations. That deeper theoretical commitment, the Gibsonian bit, largely got dropped. Roboticists took the control law and left the philosophy at the border. Whether they should have is one of the more interesting open questions in embodied AI.

3. Word Vectors → Semantic Memory

Mikolov et al. (2013): word2vec

The original motivation was engineering pragmatism. Mikolov and colleagues wanted an efficient way to learn continuous vector representations of words from very large text corpora. Word2vec was fast, scalable, and produced vectors with eerily nice algebraic properties — the "king − man + woman ≈ queen" party trick that launched a thousand conference talks.

It was, at the time, an NLP methods paper. A good one, but a methods paper.

Cognitive scientists then performed one of the more audacious heists in recent intellectual history. They took word2vec-style distributional vectors and started using them as serious models of how meaning might be organised in human minds and brains. Not as metaphors. As actual computational models that generate testable predictions.

And it worked. Reviews of semantic memory now treat distributional semantic models as major theoretical tools alongside older approaches like feature-based or taxonomic models. EEG and fMRI work has shown that word2vec-style vectors can account for significant variance in brain responses to isolated words. The geometric structure of the embedding space — which dimensions cluster which concepts, where the analogies live — turns out to have non-trivial correspondence with the structure of neural semantic representations.

That is gloriously odd. A fast NLP embedding trick, built for pragmatic engineering reasons, becomes part of the scientific argument about how meaning might be organised in biological brains.

What the receiving field improved: Theoretical interpretation. Mikolov et al. weren't making claims about cognition. Cognitive scientists asked the harder question: why do these vectors work as models of human semantic processing? That question led to deeper connections between distributional statistics, predictive processing, and theories of conceptual representation.

What got lost in transit: The original modesty. Word2vec was presented as an approximation — a useful one, but not a theory. Some of the cognitive science uptake arguably over-interpreted the correspondence, treating geometric similarity in embedding space as more cognitively meaningful than the training objective strictly licenses. The vectors are good models of something in the brain. Whether that something is semantic memory per se, or a statistical shadow of it, remains an active argument.

4. Fake Images → Scientific Instruments

Goodfellow et al. (2014): Generative Adversarial Networks

GANs were introduced as a generative-modelling trick: train a generator and a discriminator in adversarial tension until the generator's samples become indistinguishable from real data. The original paper was a machine learning contribution, and its immediate impact was in image synthesis, data augmentation, and the broader generative modelling arms race.

Cognitive scientists, meanwhile, had been suffering from a very old methodological headache. Experimental stimuli are a nightmare. You can use photographs, which are ecologically valid but uncontrollable — every real face differs on a hundred dimensions simultaneously. Or you can use simplified, parametrically controlled stimuli, which are experimentally tidy but about as naturalistic as a line drawing of a house. Ecological validity and experimental control lived in different postcodes and rarely visited each other.

GANs solved this.

Researchers now use GAN-generated stimuli — plausible-but-nonexistent faces, objects, scenes — to study perception, categorisation, novelty detection, and aesthetic judgment. The generated images are realistic enough to engage natural visual processing, but they can be parametrically varied along any dimension the experimenter cares about. You can generate a face that differs from another face only in age, or only in emotional expression, or only in some abstract latent dimension that the network learned on its own.

The punchline is excellent: fake images became scientific instruments.

What the receiving field improved: Methodological sophistication. Cognitive scientists didn't just use GANs as a stimulus factory — they developed frameworks for validating that GAN-generated stimuli actually engage the same perceptual processes as real stimuli, and for mapping the latent space onto psychologically meaningful dimensions.

What got lost in transit: The adversarial dynamics. The cognitive science uptake focused almost entirely on the generator side — the ability to produce realistic samples — and largely ignored the discriminator. Which is a shame, because the adversarial structure itself has theoretical potential. Gershman's "Generative Adversarial Brain" (2019) makes exactly this argument: that GAN-like adversarial dynamics might be a useful framework for understanding probabilistic computation in the brain, with generator and discriminator circuits in productive tension. That idea is wilder, less proven, and considerably more interesting.

The Pattern

What connects these four heists isn't just cross-pollination. It's the specific shape of the theft.

In each case, the original paper solved a problem that was local to its field. Smith & Yu were explaining infant behaviour. Lee was formalising optic flow. Mikolov was building better word vectors. Goodfellow was advancing generative modelling. None of them were writing interdisciplinary manifestos.

In each case, the migration happened because someone in the receiving field recognised a structural similarity between their problem and the original one — a similarity that was invisible from inside either field alone. The connection was not between the topics (babies and robots, embeddings and brains) but between the computational logic underneath.

And in each case, something was gained and something was lost in the crossing. The receiving field almost always improved the technique — made it faster, more general, more robust, more testable. But it also almost always dropped the theoretical context that made the original paper interesting on its own terms. The Gibsonian philosophy behind tau. The developmental scaffolding behind cross-situational learning. The modesty behind word2vec. The adversarial dynamics behind GANs.

That pattern — steal the mechanism, drop the meaning — might be the most honest description of how interdisciplinary work actually happens. Not through grand synthesis, but through selective, slightly ruthless, highly productive theft.

The question for the next decade is whether the fields can get better at stealing the meaning too. The mechanisms are flowing freely now. The theories are still getting left at the border.