Key Takeaways
- Pretrained AI models excel at mimicking behavior but were engineered for tasks like word prediction, not to reflect brain biology.
- High “brain scores” can be misleading because models may succeed via superficial correlations rather than shared neural mechanisms.
- Treating AI systems as model organisms—complex, undesigned systems to be probed—allows neuroscientists to uncover their internal computations.
- Interpretability tools that localize and perturb specific model knowledge (e.g., word‑length) reveal whether brain‑like predictions stem from genuine processing or statistical tricks.
- Brain‑tuning—aligning model representations with naturalistic brain recordings—can improve a model’s fidelity to the brain and generate new hypotheses about cortical function.
- Scaling brain‑tuning remains challenging due to data demands, but efficient, cross‑participant training strategies are emerging.
The Rise of Pretrained AI as a Neuroscience Tool
In the past decade, pretrained artificial‑intelligence (AI) models have become strikingly adept at mimicking human behavior and brain activity, prompting a “gold rush” in neuroscience. Researchers now routinely plug these models into experiments as computational stand‑ins for human cognition. Yet, as the article points out, “These models were not built to explain the brain. They were designed as engineering tools, trained to solve practical problems such as predicting the next word in a sentence.” Their strengths lie in pattern recognition, not in embodying the anatomical or evolutionary constraints that shape neural circuits.
What a “Brain Score” Really Measures
A common strategy for evaluating AI–brain similarity is to compute a brain score—the degree to which activity inside an AI model predicts recorded neural activity. Higher scores are often interpreted as evidence that the model is more brain‑like. However, the authors warn that such scores can be deceptive: modern AI models harbor rich, high‑dimensional representations that capture many input facets simultaneously. When these representations are used to predict brain activity, they may succeed for reasons unrelated to shared mechanisms, producing inflated scores that masquerade as biological insight.
The Statistical Trap of Correlated Features
The article provides a concrete illustration of this pitfall: text‑based language models, which have never processed audio, can nonetheless predict activity in auditory‑cortex regions that track speech sounds. “This isn’t magic; it’s a statistical trap. In natural language, the number of letters in a written word often correlates with the number of phonemes (sounds) in the spoken version. The auditory cortex tracks the sounds; the text‑based language model tracks the letters. Because these two are linked in the real world, the language model looks like it’s a good computational model of the auditory cortex, when it’s actually just counting characters.” The correlation between orthographic length and phonological length creates a misleading impression of neural relevance.
Why Interpretability Is Essential
To avoid being led astray by such hidden correlations, neuroscientists must treat pretrained AI models not as finished brain models but as model organisms—complex systems whose inner workings are unknown and must be discovered. The authors argue, “Like a mouse or fruit fly, they are complex systems that we did not build to test a specific neuroscientific theory. Before we can use them to learn about cognition, we must first discover what computations they perform and determine how those computations relate to the brain.” Only by probing the model’s internal representations can we test whether observed brain‑like predictions arise from genuine computational alignment.
A Case Study: Perturbing Word‑Length Knowledge
In their own work, the researchers applied an interpretability technique to isolate and perturb a language model’s knowledge of word length. After this manipulation, “the ability to predict the auditory cortex immediately vanished.” This result validated their suspicion that the model was not “listening” to speech but merely exploiting the orthographic‑phonological correlation. The quote underscores the causal power of interpretability: “The language model wasn’t actually ‘listening’; it was just exploiting a correlation.” By removing the spurious feature, the model’s brain‑like prediction disappeared, confirming that the original similarity was superficial.
From Model Organisms to Brain‑Tuned Systems
Beyond diagnosing misalignments, the article proposes a proactive strategy: brain‑tuning. This process refines a pretrained model’s internal representations by aligning them with brain recordings gathered during naturalistic tasks—such as listening to an audiobook or watching a movie—that simultaneously engage perception, language, memory, and prediction. Unlike earlier approaches that used narrow, object‑recognition data, brain‑tuning captures the brain’s rich, multimodal dynamics. The goal is not merely to boost neural prediction scores but to shape the model into a more faithful organism whose computations mirror those of the human cortex.
Evidence That Brain‑Tuning Works
Early experiments show promising results. When a language model is brain‑tuned on auditory data, it does not just improve at predicting the specific training set; it becomes a “better general listener.” The tuned models can forecast brain activity for entirely new listeners and novel narratives, picking up features that go beyond simple word length. Moreover, “these tuned models start to process features of speech that we haven’t even identified yet.” By reverse‑engineering these emergent sensitivities, researchers can generate fresh hypotheses about how the auditory cortex extracts speech information, turning the AI model into a source of neuroscientific insight rather than a black‑box mimic.
Matching Hierarchical Brain Processing
An additional benefit of brain‑tuning is that the model’s representation space begins to resemble the hierarchical organization seen in the brain. Pretrained AI models often develop flat or entangled feature spaces, whereas brain‑tuned models exhibit layered patterns akin to the cortical hierarchy—from low‑level acoustic cues to higher‑level linguistic meaning. This structural alignment strengthens the claim that the tuned model is not merely correlating with brain activity but is emulating the brain’s computational architecture.
Challenges Ahead: Data, Mapping, and Efficiency
The authors acknowledge substantial hurdles. Mapping AI computations onto neural mechanisms remains an open problem because “AI and brains are built differently.” Furthermore, brain‑tuning is data‑hungry: refining a large model demands extensive brain recordings, far exceeding what a single study typically yields. To mitigate this, the team is experimenting with efficient training schemes that enable tuning across multiple participants using smaller datasets. Scaling these methods will be essential if brain‑tuned models are to become widely usable tools in cognitive neuroscience.
Toward a Mechanistic Understanding of Cognition
Ultimately, shifting the perspective from “AI as a finished brain model” to “AI as a model organism to be perturbed, tuned, and reverse‑engineered” promises a deeper grasp of neural mechanisms. By leveraging interpretability to uncover hidden correlations, employing brain‑tuning to align models with natural brain dynamics, and iteratively testing causal hypotheses, neuroscientists can move beyond descriptive correlations toward genuine mechanistic explanations of how the brain perceives, understands, and predicts the world. As the article concludes, this approach brings us closer to cognitive neuroscience that “doesn’t just describe the brain but truly understands its mechanics.”