Neuroscience Citations in AI Papers Grow in Number, Shrink in Share

The absolute number of AI papers citing neuroscience has climbed steeply since 2015, while the fraction of AI papers citing neuroscience has stayed low and perhaps declined. That contrast sits at the center of Renato Duarte's citation-pattern analysis of the neuroscience–AI interface, and it frames a practical question for researchers deciding where to invest their time after the post-2022 AI boom.

Duarte examined cross-field citations between neuroscience and artificial intelligence and found that AI has drawn mainly on loose, abstract concepts from neuroscience rather than on specific findings about neural mechanisms. A responding commentator, writing after Duarte's piece, stated that this observation "matches my own experience."

What do the citation numbers actually show?

The commentary's most cited plot contrasts two measures:

• The absolute number of AI papers citing neuroscience, which began to climb steeply after 2015.
• The fraction of AI papers citing neuroscience, which remained low and may have dipped over the same period.

Both directions of cross-citation — neuro-citing-AI and AI-citing-neuro — grew in absolute terms after 2015. The fraction of neuroscience papers citing AI also increased. Only the fractional share of AI papers citing neuroscience failed to rise.

The takeaway: neuroscience-influenced AI work is growing in volume even as it becomes a smaller percentage of a rapidly expanding AI literature.

Is neuroscience's influence on AI rising or falling?

The commentator lays out both readings. On one hand, the declining fraction of neuro citations suggests neuroscience was not a dominant driver of the AI boom compared with other factors. On the other hand, the rising absolute number of neuro-inspired AI papers "still represents a substantial footprint," even at a shrinking percentage of the overall literature.

This distinction carries strategic weight for researchers at the NeuroAI interface. The commentary poses two competing strategies for the post-2022 landscape:

• Treat any possibility of influencing AI, even a fringe one, as a competitive route to maximal impact, given AI's scale.
• Treat AI instead as a toolkit and find lower-hanging fruit in applying it to the hard problem of understanding the brain.

The commentator does not resolve the choice, but frames it as the core question for the field's current moment.

Where do the parallels between brains and current AI actually lie?

The commentary identifies design choices in the current generation of AI tools that parallel purported properties of the brain:

• Fast activity-based mechanisms for reasoning.
• Geometric codes that expose relative positions.
• Planning through mental simulation.

The commentator attributes these parallels to a paper co-authored with Chris Summerfield, published earlier this month. The claim is careful: these are parallels between engineering choices and purported brain properties, not evidence that AI systems implement neural mechanisms.

For readers of Duarte's underlying argument, the point matters. If AI design borrows abstract concepts rather than specific neurophysiological findings, then citation counts may understate or misrepresent the depth of the interdisciplinary exchange.

A call for new tools and results

Duarte's piece closes with a call for new results and tools to refresh the interdisciplinary relationship between the two fields. The commentator endorses that call, positioning it as the constructive response to a citation picture that is otherwise ambiguous in its message.

A companion comment by Tony Zador, "Citation graphs may underestimate neuroscience's influence on AI," extends the debate from the measurement side.

The open question for NeuroAI researchers is whether the next cycle of cross-field influence will come from neuroscience shaping AI systems, or from AI accelerating the mechanistic understanding of the brain.