Citation Graphs Undercount Neuroscience's Role in AI, Commentary Argues

Roughly two-thirds of references in contemporary AI conference papers are five years old or younger, with a median age of just 3 years — against a median of 8 years for neuroscience journal papers, with citation tails stretching past 20 years. That gap sits at the center of a new commentary arguing that citation analysis systematically underestimates how much modern artificial intelligence owes neuroscience.

The piece responds to Renato Duarte's citation analysis of cross-field traffic between neuroscience and AI. Duarte's map of how the two fields cite each other is valuable, the commentator writes, but it cannot support the claim that "the astonishing recent achievements in AI owe neuroscience surprisingly little in their actual machinery and operating principles." Citations, the argument goes, are simply the wrong instrument for measuring intellectual debt.

Why citation counts distort influence

Modern AI is increasingly an engineering discipline. A handful of conceptual innovations can generate enormous downstream literatures devoted to optimization, scaling, efficiency and applications — papers that cite their immediate technical predecessors rather than the deeper ancestry of the ideas they use. If one organizing idea from neuroscience generates 10,000 engineering papers, citation counts will make the original contribution look negligible, even if it redirected the field.

The transformer architecture underlying large language models such as ChatGPT illustrates the point. The lineage runs from recurrent neural networks, which framed sequence processing around a learned internal state evolving through time. RNNs in turn drew on the long-standing neuroscience view of recurrent circuits as substrates for temporal integration, memory and persistent activity. Attention relaxed the requirement that information pass through a single evolving state. The transformer then discarded recurrence entirely once arbitrary positions could interact through attention — an architecture that parallelized and scaled far more effectively on GPU hardware.

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 tied to a single canonical citation. The same problem afflicts attention itself, a concept developed over decades in psychology and neuroscience before becoming a particular mathematical operation in machine learning. By the time the famed "Attention is all you need" paper appeared, the natural proximate citations were earlier machine-learning papers.

What is 'obliteration by incorporation'?

A citation graph records the last few handoffs in an intellectual lineage far better than the lineage itself. The commentator points to a pattern the sociologist Robert Merton named "obliteration by incorporation": successful ideas become so thoroughly absorbed into a field's common vocabulary that citing their origin becomes unnecessary.

The analogies are familiar to any reader of the scientific literature. 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 became foundational.

The problem extends to recent borrowings, not just old ones. The commentary cites Anthropic's "Golden Gate Claude" — a notable advance in AI interpretability made vivid by the recent Hugging Face incident that pushed AI safety into public view — as owing an intellectual debt to the past decade of neuroscience work on mixed selectivity and population representations. Yet that lineage is largely absent from the paper's citations; the neuroscience papers it does cite are mostly decades older.

Does the people-trail tell a different story?

The commentary proposes an alternative, less quantitative method: tracing the people who carried ideas between fields. John Hopfield, who won a Nobel Prize for foundational work on neural networks, developed his ideas 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. The influence continues into the next generation: the founders of two of the three largest AI labs, Demis Hassabis of DeepMind, a Nobel laureate for AlphaFold, and Dario Amodei of Anthropic, focused on neuroscience in their doctoral research.

The commentator acknowledges the spot check of citation ages needs systematic follow-up. Duarte's analysis remains a valuable map of current cross-field traffic — but answering how much modern AI owes neuroscience requires intellectual history, tracing ideas as they migrate between fields, change names, become formalized and eventually disappear into the technical substrate.