Everyone wants to build AI that behaves like the human brain — but here’s the problem: even two real brains (human or mouse) don’t respond exactly the same way to the same stimulus.
So how do you fairly measure whether an AI is “brain-like” at all?
A new paper by Yamins and colleagues proposes a simple but brilliant fix.
Instead of comparing AI and brain responses directly, they first ask:
“What kind of transformation does it take to turn the neural signals of one brain into another?”
Then — they apply that same transformation to an AI model.
If the AI can be “disguised” as a normal brain using the same kind of mapping that connects two real brains, it’s not just mimicking — it’s genuinely brain-like.
📊 How it works:
- Researchers collected data from mice (electrophysiology) and humans (fMRI).
- They learned how to “translate” one individual’s neural responses into another’s — a process called the Inter-Animal Transform Class (IATC).
- They then applied the same transformation to AI activations, checking whether the model could “blend in” like a biological brain.
💡 Key Findings:
- AI models could be “masked” as real brains using the same transformations that link biological ones.
- TDANNs (Topographic Deep Artificial Neural Networks) were the most brain-like — their internal structures mirrored the cortical maps found in real brains.
- The method worked across species — from mouse to human.
- It captured subtle differences between model architectures and even between activation functions (ReLU vs GELU).
- Most importantly — it offers a fair comparison: no more unrealistic one-to-one matches, just consistent transformations.
Before, scientists expected AIs and brains to align perfectly — which even real brains don’t do.
Now, the rule is simple: if an AI can “blend in” among real brains using the same neural transformation tricks, it earns the label brain-like.
For the first time, researchers have a fair, testable way to evaluate how close AI really comes to the human mind.
🧩 AI doesn’t have to be a clone of the brain — just another individual in the population.
📄 Paper: “Towards Fair Comparisons of Brains and Models” — arxiv.org/abs/2510.02523
