In May 2026, a paper from the Weizmann Institute and MIT dropped a bombshell: after running causal tests on 260 visual concepts, the authors found that more than 70% of concept-specific brain regions previously identified by activation methods were false positives.
What does this mean? For the past two or three decades, the colorful brain maps in countless fMRI papers—hot spots labeled "this region encodes concept X"—may have been marking, seven times out of ten, a region merely responding to some incidental co-occurring feature: color, background, posture, splashing water—not encoding the concept itself.
If this sounds like the old cliché "correlation does not imply causation," that's exactly right. The problem is that brain science has been treating correlation as causation for decades.
Where the Problem Lies: Activation ≠ Encoding
The basic logic of fMRI imaging is simple: show subjects a set of images, record blood-oxygen responses (BOLD signals) across brain regions, and ask "which region responds more strongly to 'faces' than to other categories?"
This approach—the category contrast method—has been used since the 1990s. It discovered classic functional regions such as the FFA (fusiform face area), EBA (extrastriate body area), and PPA (parahippocampal place area).
But there is a fatal flaw.
Suppose you show subjects images of "surfing" and find that some brain region responds significantly more to surfing images than to other images. Excitedly, you announce: this region encodes the "surfing" concept!
Wait. What else is in a surfing image? Ocean water. Human bodies. Standing posture. Blue sky. Motion blur. Is the region responding to "surfing" or to "water + human body + stance"?
The traditional method cannot distinguish these cases. It only compares average responses between "with surfing" and "without surfing"—but the "without surfing" control group usually lacks water, bodies, and stances too. You have no idea which feature the region is actually tracking.
It's like noticing a thermometer reads high when water boils and declaring "the thermometer encodes boiling"—when it's actually just tracking temperature, which happens to accompany boiling.
BrainCause: Adding a Causal Engine to Brain Science
BrainCause's core idea: construct three types of stimuli per target concept to fully separate the concept itself from its co-occurring associative features.
- Positive samples: images containing the target concept, e.g., "surfing"—a person on a surfboard with waves.
- Semantic negatives: semantically related images that lack the target concept, e.g., "fishing" (person + water, no surfing), "beach" (water + sand, no surfing), "waves" (water + motion, no surfing). These preserve most associative features while removing the concept itself.
- Counterfactual samples: the target concept removed or replaced while everything else stays as unchanged as possible, e.g., erasing the surfer and leaving waves and sky, or swapping the surfboard for a floating plank.
- 73.4% of the 260 concept regions found by activation methods fail causal verification—they respond to semantic negatives and counterfactual edits as strongly as to positives, meaning they track co-occurring associative features, not the concepts themselves.
- BrainCause's causal ranking reduces 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 within FFA and EBA/FBA—not a blurry "body" representation but fine-grained part distinctions
- Text types: handwriting, signage, and logos each have preferred subregions within VWFA and OWFA
- Tools: localized near body and action-related regions—consistent with "tools as extensions of the body"
- Animal faces: falling within known face-selective regions (FFA, OFA)
- Social interaction: causally verified specific representations
This design mirrors randomized controlled trials (RCTs): positives are the treatment group, semantic negatives control for semantic confounds, and counterfactuals control for visual confounds. A region "causally" encodes a concept only if its response to positives is simultaneously significantly higher than to both negative types.
A Fully Automated Pipeline: From Datasets to Follow-Up Experiments
BrainCause is not just a scoring method—it is a fully automated framework:
1. Causal dataset construction: Gemma-3-27B generates diverse prompts, FLUX.2 generates images, and Qwen3-VL-8B verifies whether images truly contain/lack the target concept. Each concept gets 200 positives, 100 semantic negatives, and 100 counterfactual edits.
2. Representation search: for each voxel, three scores are computed—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 become candidates.
3. Final judgment and follow-up experiment design: causal evidence strength and measurement coverage are evaluated. If coverage is insufficient (e.g., too few "dog" images in NSD), BrainCause automatically generates supplementary stimuli, directly guiding the next round of fMRI experiments.
This "auto-completion" feature is especially practical. Traditional fMRI experiment design is manual—researchers pick images from experience and often miss key control conditions. BrainCause can tell you: "Your current dataset lacks semantic negatives for 'surfing'; here are suggested images…"
The 70% False Positives: What the Numbers Mean
BrainCause ran large-scale causal tests on 260 visual concepts, compared against MindSimulator, the strongest prior activation method:
The "surfing" example is telling: MindSimulator claimed to find a surfing-specific region, but causal testing showed it responds to "fishing," "beach," and "waves" images just as strongly. It wasn't encoding "surfing"—it was encoding "water + human body."
New Findings After Causal Verification
The good news: BrainCause doesn't just reject false positives—it also finds true ones.
It accurately reproduces the classic functional regions: FFA (faces), EBA (bodies), PPA (scenes), VWFA (visual word form)—all survive causal testing, confirming genuine causal specificity.
More importantly, it discovers many causally verified fine-grained concept representations:
BrainCause's Limitations
BrainCause depends on current language and vision models to generate and verify stimuli. If the language model proposes poor "counter-concepts" (e.g., for "surfing" thinking only of "beach" but not "skateboarding"), or the image-editing model cannot cleanly remove the target concept, the causal tests may miss some confounds.
BrainCause's own false positives confirm this: they occur mainly for broad visual attributes like "sky," "reflection," and "lighting contrast"—too low-level and pervasive to fully control with semantic negatives and counterfactual edits.
The Deeper Implication
The paper's core message is not merely "70% of old findings are false." It points to a more fundamental question: in brain science, what standard are we actually using to claim "the brain encodes concept X"?
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 goes off during a fire—it's tracking 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 lacking X, after controlling for all associative features that co-occur with X.
This standard is stricter, but more honest. It means many concept representations we thought we'd "found" need re-verification. And it means future brain science can no longer settle for drawing an activation map and announcing a discovery—you must prove you've found the fire, not the smoke.
---
Paper: From Activation to Causality: Discovery of Causal Visual Representations in the Human Brain
Institutions: Weizmann Institute of Science · MIT
Dataset: Natural Scenes Dataset (NSD, 7T fMRI, ~10,000 natural images per subject across 8 subjects)