In May 2026, a paper from the Weizmann Institute and MIT dropped a bomb: after running causal tests on 260 visual concepts, it found that over 70% of concept-specific brain regions previously identified by activation-based methods were false positives.
What does this mean? It means that across the last two to three decades, on those colorful brain maps in countless fMRI papers, the hotspots labeled "this region encodes concept X" had a roughly 70% chance of merely reflecting responses to some accidentally co-occurring feature—color, background, posture, water splashes—rather than encoding the concept itself.
If this sounds like the old cliché "correlation is not causation," you're right. The problem is that neuroscience has been treating correlation as causation for decades.
The problem: activation ≠ encoding
The basic logic of fMRI brain imaging is simple: show subjects a set of images, record the blood-oxygen response (BOLD signal) of each brain region, and ask, "Which region responds more strongly to 'faces' than to other categories?"
This method, called category contrast, has been used since the 1990s. It produced classic functional regions like the FFA (fusiform face area), EBA (extrastriate body area), and PPA (parahippocampal place area).
But there is a fatal flaw.
Suppose you show subjects "surfing" images and find that some brain region responds significantly more to surfing images than to others. You excitedly announce: this region encodes the "surfing" concept!
Wait. What else is in a surfing image? Ocean water. Human bodies. A standing posture. Blue sky. Motion blur. Is the region responding to "surfing," or to "water + human body + standing"?
Traditional methods cannot distinguish these cases. They only compare average responses between "with surfing" and "without surfing"—but the "without surfing" control images usually lack water, bodies, and posture too. You have no idea which feature the region is tracking.
It's like noticing a thermometer reads high during boiling and declaring "the thermometer encodes boiling"—when it's really just tracking temperature, which happens to coincide with boiling.
BrainCause: a causal engine for brain science
BrainCause's core idea: build three types of stimuli for each target concept to fully separate the concept itself from co-occurring correlated features.
- Positive samples: images containing the target concept, e.g., "surfing"—a person on a surfboard with waves.
- Semantic negatives: semantically related images that do not contain the concept, e.g., "fishing" (person + water, no surfing), "beach" (water + sand, no surfing), "waves" (water + motion, no surfing). These retain most correlated features while lacking the concept itself.
- Counterfactual samples: the target concept removed or replaced while everything else stays as unchanged as possible—erase the surfer, leaving waves and sky; or swap the surfboard for a floating plank.
- Of the 260 concept regions found by activation methods, 73.4% failed causal validation—they responded to semantic negatives and counterfactual edits just as strongly as to positives, meaning they tracked correlated features, not the concepts themselves.
- BrainCause's causal ranking cut the false positive rate from 73.4% to 23%, while raising the true positive rate from 26.6% to 38.7%.
- Body parts: faces, hands, and legs show distinct distribution patterns across FFA and EBA/FBA—fine-grained part distinctions, not a blurry "body" representation
- Text types: handwriting, signs, and logos each have preferred subregions within VWFA and OWFA
- Tools: localized near body and action-related areas—consistent with the intuition that tools are extensions of the body
- Animal faces: fall within known face-selective regions (FFA, OFA)
- Social interaction: has a causally validated specific representation
This design mirrors a randomized controlled trial (RCT): positives are the treatment group, semantic negatives control semantic confounds, and counterfactuals control visual confounds. Only when a region's response to positives is significantly higher than to both types of negatives can it be said to *causally* encode the concept.
Automated end-to-end pipeline
BrainCause is a fully automated framework, not just a scoring method:
1. Causal dataset construction: Gemma-3-27B generates diverse prompts, FLUX.2 generates images, and Qwen3-VL-8B verifies whether images truly contain or exclude the concept. Each concept gets 200 positives, 100 semantic negatives, and 100 counterfactual edits. 2. Representation search: each voxel receives three scores—activation score (response to positives), semantic causal score (positives vs. the hardest semantic negatives), and counterfactual causal score (positives vs. the hardest counterfactual edits). Only voxels high on all three enter candidate regions. 3. Final adjudication and follow-up design: it evaluates causal evidence and measurement coverage. If coverage is insufficient (e.g., too few "dog" images in the NSD dataset), BrainCause automatically generates the supplementary stimuli needed for the next fMRI experiment.
This "auto-completion" feature is especially practical. Traditional fMRI design is manual—researchers pick images by experience and often miss key control conditions. BrainCause can tell you: "Your dataset lacks semantic negatives for 'surfing'; add these images..."
The 73.4% figure
BrainCause ran large-scale causal tests on 260 visual concepts, comparing against MindSimulator, the strongest prior activation-based method:
The "surfing" example is telling: MindSimulator claimed a surfing-specific region, but causal testing showed it responded to "fishing," "beach," and "waves" images just as strongly. It wasn't encoding surfing—it was encoding the combination of "water + human body."
New findings after causal validation
The good news: BrainCause didn't just reject false positives—it found true ones.
It accurately reproduced the classic functional areas: FFA, EBA, PPA, and VWFA (visual word form area) all survived causal testing, confirming genuine causal specificity.
More importantly, it discovered many causally validated fine-grained concept representations:
Limitations
BrainCause relies on current language and vision models to generate and verify stimuli. If the language model proposes poor "counter-concepts" (e.g., for "surfing" it thinks of "beach" but not "skateboard"), or the image editing model cannot cleanly remove the target concept, the causal test may miss confounds.
BrainCause's own false positives confirm this: they concentrate on broad visual attributes like "sky," "reflections," and "lighting contrast"—too low-level and pervasive to be fully controlled by semantic negatives and counterfactual edits.
The deeper implication
The paper's core message isn't just "70% of old findings are false." It points to a more fundamental question: in neuroscience, what standard do we use to claim "the brain encodes a concept"?
If the standard is merely "stronger activation," we can never distinguish encoding from association. You can't say a smoke alarm "encodes fire" just because it rings during a fire—it tracks smoke, which happens to co-occur with fire.
BrainCause's proposed standard: a brain region encodes concept X if and only if its response to images containing X remains significantly higher than to images without X, after controlling for all features that co-occur with X.
This standard is stricter, but more honest. It means many concept representations we thought we'd "found" need revalidation. And it means future neuroscience can no longer be satisfied with drawing an activation map and announcing a discovery—you must prove that what you found is not the smoke, but the fire.
---
Paper: From Activation to Causality: Discovery of Causal Visual Representations in the Human Brain
Project page: see the paper link
Institutions: Weizmann Institute of Science · MIT
Dataset: NSD (Natural Scenes Dataset; 7T fMRI, ~10,000 natural images per subject across 8 subjects)