458 Million Papers Later: The Neuroscience–AI Bridge Is Real but Narrow

A citation analysis of roughly 458 million publications and the 2.9 billion citation links among them finds that the neuroscience–artificial intelligence connection is real and accelerating — but also narrow, tool-heavy, and concentrated in a small set of institutions. No landmark paper published after 2003 has become essential reading shared by both fields.

The analysis, drawn from the open publication database OpenAlex and covering 1995 to 2025, was designed to answer an empirical question: not whether neuroscience and AI should fuel each other, but whether they actually do, and how knowledge moves between them.

What the numbers show

Corrected for field size, neuroscience has influenced AI more than AI has influenced neuroscience over the past three decades. The chance-corrected likelihood of an AI paper citing neuroscience work is higher than the reverse across the entire window examined.

The driver, the analysis finds, is a small fixed canon of classical neuroscience papers — the conceptual basis for artificial neural networks and deep learning. Bi and Poo's 1998 work on spike-timing-dependent plasticity and Izhikevich's 2003 spiking neuron models still anchor the bridge, alongside work on visual attention and saliency, object recognition, receptive fields and sparse coding, working memory, and short-term plasticity. The most recent of these bridge works dates from 2003.

The rate of exchange, however, is shifting fast. Neuroscience's citations of AI rose steadily from 2006 and accelerated between 2016 and 2019, tracking the widespread adoption of machine-learning tools for neural data analysis. AI's citations of neuroscience declined from 2006 to 2019, then climbed again — with 2020 to 2025 showing the steepest slope in the series, consistent with the timing of the NeuroAI manifestos.

Of the neuroscience that AI papers now cite, about 90 percent comes from cognitive neuroscience and less than 5 percent from cellular and molecular work.

How strong are the caveats?

Two cautions matter. OpenAlex data from 2024 and 2025 are unstable because the database is mid-rebuild, so recent magnitudes cannot be trusted, though the upturn predates those years and is not an artifact. And the steep post-2020 rise may partly reflect neuroscience shrinking as a share of an exploding AI literature rather than AI citing it more often. Between 2020 and 2023, the raw share of AI citations going to neuroscience rose substantially, but more modestly than the odds ratio suggests.

The author chose the odds ratio — how often one field cites the other relative to random expectation given field size — as the more honest view. On raw shares, the lines cross in 2018, when AI's influence on neuroscience takes the lead.

Which papers, authors, and institutions hold the bridge?

The analysis ranked bridge papers by the harmonic mean of citations from each field — a metric that punishes imbalance, so a paper cited 1,000 times by AI and twice by neuroscience scores low, while one cited 100 times by each side scores high. Roughly 59,000 papers with at least five citations from each side were considered.

About half of the strongest bridges are not ideas at all, but tools: general-purpose statistical and machine-learning software cited routinely on both sides, including the BraTS tumor-segmentation challenge, the lme4 mixed-effects package, SciPy, ROC analysis, and XGBoost.

The author ranking mixes AI-leaning researchers — Yoshua Bengio, Geoffrey Hinton, Antonio Torralba — with computational neuroscientists including Karl Friston, Peter Dayan, Wulfram Gerstner, Henry Markram, Christof Koch, and Aude Oliva. Rank is earned by simultaneous citation pull from both fields, not eminence in one.

The institutional skew is the sharpest finding. Two industry labs, Alphabet (Google plus DeepMind) and Microsoft, top the ranking. MIT, UC Berkeley, Harvard, University College London, and Stanford follow. The top 100 institutions absorb nearly 40 percent of all bridge weight; the top 100 authors account for under 7 percent. The bridge, in short, is held up by institutional resources and environments, not by individual mobility.

What changes for the field?

The essay's argument is direct: inspiration alone will no longer keep neuroscience influential. AI's recent gains owe the brain little in actual machinery. For relevance, neuroscience must deliver results AI can use on its own problems — tools to dissect AI's internal workings, and routes to more energy- and data-efficient architectures.

The stakes are concrete, not metaphorical. The human brain runs on roughly 20 watts, learns from a handful of examples, and generalizes out of distribution; energy efficiency, sample efficiency, and robustness are engineering targets frontier labs are already chasing. The essay also warns that institutional concentration breeds narrow consensus, and that the oldest, hardest problems require wider participation.

All code, data, and methods — including field assignments, the citation-flow analysis, and a hand verification of every affiliation — are available in a public GitHub repository. Whether the post-2020 uptick becomes a sustained renaissance, the author concludes, depends on whether the community gives AI something newer to cite than the works it produced 30 years ago.