Citation graphs may underestimate neuroscience’s influence on AI

Renato Duarte’s citation analysis provides an interesting map of how neuroscience and artificial intelligence (AI) cite each other, but I think it overstates what those patterns can tell us about intellectual influence. In particular, the claim that “the astonishing recent achievements in AI owe neuroscience surprisingly little in their actual machinery and operating principles” goes well beyond what citation analysis can establish. Citations are simply not a good measure of that kind of intellectual debt.

Modern AI is increasingly an engineering discipline. A small number of conceptual innovations can generate enormous downstream literature devoted to optimization, scaling, efficiency and new applications. Those papers naturally cite their immediate technical predecessors, not the deeper intellectual ancestry of the ideas they are using. If a single key idea—from neuroscience or elsewhere—contributes one organizing idea that then generates 10,000 engineering papers, citation counts will typically make the original contribution look negligible, even if it helped redirect the field.

The history of the transformer—the architecture underlying modern large language models such as ChatGPT—provides an illustrative example. The transformer can be traced back to recurrent neural networks (RNNs), which framed sequence processing around a learned internal state that evolves through time. RNNs in turn drew on the long-standing neuroscience view of recurrent neural circuits as substrates for temporal integration, memory and persistent activity. Attention then relaxed the requirement that all relevant information pass through a single evolving state. The transformer took the next step: Once arbitrary positions could interact through attention, recurrence itself could be discarded. What began in part as a way of improving sequence models became an architecture that could be parallelized and scaled far more effectively on the GPU, the specific hardware that has come to dominate modern AI.

But what neuroscience paper should an RNN paper cite for recurrent computation? There is no obvious answer. Recurrent circuitry is textbook neuroscience, not a discovery associated with a single canonical citation. The same problem arises with attention. “Attention” was a broad concept developed over decades in psychology and neuroscience before being turned into a particular mathematical operation in machine learning. By the time the famed “Attention is all you need” paper appeared, the appropriate proximate citations were naturally earlier machine-learning papers on attention, not some single neuroscience paper that could stand for the intellectual tradition from which the concept emerged.

A citation graph therefore records the last few handoffs in an intellectual lineage much better than it records the lineage itself. This is not a peculiarity of neuroscience and AI. Imagine trying to estimate the contribution of mathematics to modern engineering from direct citations. Signal-processing papers do not routinely cite Joseph Fourier. Engineers using linear algebra do not cite Arthur Cayley or James Joseph Sylvester. Modern molecular biology papers rarely cite James Watson and Francis Crick. Foundational ideas often become invisible precisely because they have become foundational. Once they enter the common vocabulary of a field, citing their origin becomes unnecessary. The sociologist Robert Merton called this “obliteration by incorporation”: Successful ideas can become so thoroughly absorbed that their sources disappear from the citation record. A method based on direct citation therefore has a systematic tendency to underestimate deep, old, successful influences.

The short citation memory of machine learning reinforces this interpretation. AI, especially conference-based machine-learning literature, appears to cite mainly its immediate antecedents; in a small spot check of contemporary AI conference papers, roughly two-thirds of references were 5 years old or younger, with a median age of only 3 years. By contrast, neuroscience journal papers had a median of 8 years, with a much longer tail extending back 20 years or more. A more systematic analysis is needed, but it appears these citation conventions strongly favor recent technical antecedents over deeper intellectual ones.

Obliteration by incorporation shows how the impact of foundational ideas can fade from the citation record. But the same problem can arise when newer ideas are borrowed from neuroscience. The recent Hugging Face incident made the stakes of the AI safety debate vivid to the public. Some of the most important advances aimed at addressing this problem, such as Anthropic’s “Golden Gate Claude,” owe an intellectual debt to neuroscience work on mixed selectivity and population representations over the past decade. Yet that lineage is largely absent from the paper’s citations; the neuroscience papers it does cite are mostly decades older.

These considerations suggest that an alternative, albeit less quantitative, way to trace this history is through the people who carried ideas between the fields. It is hardly accidental that many of the researchers responsible for foundational developments in AI were deeply engaged with neuroscience and psychology. John Hopfield, who won a Nobel Prize for his foundational work on neural networks, developed his ideas about neural computation in close dialogue with neuroscientists studying biological neural systems; the “Godfathers of AI”—Geoffrey Hinton (who shared the Nobel with Hopfield), Yann LeCun and Yoshua Bengio—did likewise. That influence extends to the next generation. The founders of two of the three largest AI labs, Demis Hassabis of DeepMind, who won a Nobel for his work on AlphaFold, and Dario Amodei of Anthropic, focused on neuroscience in their doctoral research.

Duarte’s analysis is a valuable map of current cross-field traffic, but it is not a measure of how much modern AI owes neuroscience. That question requires more than just bibliometrics. It requires intellectual history: tracing ideas as they migrate between fields, change names, become formalized and eventually disappear into the technical substrate.

Read Renato Duarte’s piece: Examining the bridge between neuroscience and AI

Read Kim Stachenfeld’s comment: The rising absolute number of neuro-inspired AI papers represents a substantial footprint