40% of fMRI Brain Signals Inverted: AI Training Data Crisis

Researchers at Technical University of Munich (TUM) published a study in Nature Neuroscience on December 12, 2025, revealing that approximately 40% of functional MRI (fMRI) signals contradict actual brain activity. Increased signals can correspond to reduced neuronal firing, and decreased signals can mean heightened activity—the exact opposite of assumptions that have guided three decades of neuroscience research. For developers building AI systems on brain imaging data or working with brain-computer interfaces, this is a training data crisis: your labels are 40% wrong.

The implications cascade through the entire AI/ML ecosystem. Machine learning models trained on fMRI datasets—used for diagnosing Schizophrenia, Alzheimer’s, and other brain disorders—may be overfitting to vascular noise instead of neural signals. Brain-computer interfaces achieving 82% accuracy in speech decoding might be reading blood flow patterns, not thoughts. Neural networks marketed as “brain-inspired” may be built on neuroscience that doesn’t reflect biological reality.

Brain Oxygen Extraction Breaks fMRI Assumptions

The traditional model was straightforward: neurons fire more, demand more oxygen, trigger increased blood flow, and fMRI detects the signal boost. TUM’s study, led by Dr. Samira Epp and Prof. Dr. Valentin Riedl, shattered this. Testing 40+ participants performing mental arithmetic and memory recall, they simultaneously measured fMRI signals and actual oxygen consumption using novel quantitative MRI techniques. What they found contradicts decades of established practice.

Active brain regions frequently meet energy demands by extracting oxygen more efficiently from existing blood supplies—without increasing blood flow. In roughly 40% of cases, increased neuronal activity registers as a decreased fMRI signal. Dr. Epp puts it directly: “This contradicts the long-standing assumption that increased brain activity is always accompanied by increased blood flow to meet higher oxygen demand.” The coupling neuroscientists relied on isn’t universal—it’s correct about 60% of the time.

AI Training Data Corrupted by fMRI Signal Inversion

This has immediate consequences for artificial intelligence and machine learning. fMRI datasets like the Natural Scenes Dataset (8 participants, thousands of images, 1 year of scanning) and the Human Connectome Project are foundational for training models that decode brain states, diagnose disorders, and validate neural network architectures. Publications combining fMRI and machine learning exploded from 39 in 2010 to 1,165 in 2020. That growth is now on shaky ground.

Deep learning models achieve 94.31% accuracy on Human Connectome Project task-evoked fMRI data. However, if 40% of the training labels are inverted—activity labeled “high” when it’s actually “low”—those benchmarks may represent sophisticated overfitting to vascular artifacts, not genuine understanding of brain function. For medical AI systems diagnosing psychiatric or neurological conditions from fMRI scans, a 40% label noise rate is catastrophic. Garbage in, garbage out.

Tens of Thousands of Neuroscience Studies Need Reinterpretation

fMRI has been the neuroscience gold standard since the mid-1990s. Tens of thousands of published studies on depression, Alzheimer’s, cognitive processes, and more relied on the blood flow coupling assumption. Prof. Riedl is blunt: interpretations of psychiatric and neurological conditions “must now be reassessed,” particularly for patient populations with vascular changes from aging, stroke, or disease.

The problem compounds in clinical contexts. Researchers may have been measuring vascular dysfunction—impaired blood flow regulation—and interpreting it as neuronal deficits. These require fundamentally different treatments. How many diagnoses, clinical trials, or therapeutic interventions were based on signals that meant the opposite of what scientists thought? This is a reproducibility crisis with real-world consequences.

Brain-Computer Interfaces and ‘Brain-Inspired’ AI Face Reality Check

Real-time fMRI-based brain-computer interfaces decode signals to control devices or reconstruct speech, achieving up to 82% accuracy with GPT-based decoders. If 40% of fMRI signals are inverted, these systems may be reading vascular patterns rather than neural intent. BCIs relying solely on fMRI need recalibration and validation against direct neural measurements like EEG (which measures electrical activity and isn’t affected by this blood flow issue).

“Brain-inspired” neural networks face a similar reckoning. Convolutional architectures modeled after the visual cortex, hierarchical processing systems, and representational similarity analyses comparing AI activations to fMRI patterns—all rest on neuroscience findings that may be partially wrong. Brain-inspired doesn’t mean brain-accurate. Successful AI architectures often diverge significantly from biological reality anyway, but this finding underscores the need for empirical validation over assumed neuroscience truths.

Multi-Modal Validation Essential for Neuroimaging Reliability

TUM’s solution is straightforward: don’t rely on fMRI alone. Use quantitative MRI to directly measure oxygen consumption. Combine fMRI with EEG (electrical activity), PET scans (glucose metabolism), or other modalities that aren’t subject to the same blood flow coupling failures. For AI/ML developers working with neuroimaging data, this means ensemble approaches that cross-validate findings across multiple measurement techniques.

Treat fMRI-derived labels as noisy estimates, not ground truth. Apply robust loss functions designed to handle label corruption. Question whether high accuracy on fMRI benchmarks reflects true understanding or artifact overfitting. The broader lesson: assumptions about training data quality can persist unchallenged for decades. Validating your ground truth labels isn’t paranoia—it’s engineering rigor.

Key Takeaways

• 40% of fMRI signals show an inverse relationship with actual brain activity—increased signals can mean decreased neuronal firing (TUM study, Nature Neuroscience, December 12, 2025)
• AI models trained on fMRI data may be overfitting to vascular noise rather than neural signals; 40% label corruption undermines supervised learning benchmarks
• Tens of thousands of neuroscience studies from the past 30 years may require reinterpretation, particularly clinical research on brain disorders
• Brain-computer interfaces and “brain-inspired” neural networks need validation against direct neural measurements (EEG) and empirical performance, not assumed neuroscience accuracy
• Multi-modal neuroimaging (fMRI + EEG + PET) provides robustness against single-modality failures; treat fMRI as one noisy signal among many, not ground truth