Plasticity and Language in the Anaesthetized Human Hippocampus

⬅️ Back to Articles

📝 ARTICLE INFORMATION

• Article: Plasticity and language in the anaesthetized human hippocampus
• Authors: Katlowitz, K. A., Cole, E. R., Mickiewicz, E. A., Shah, S., Franch, M., Adkinson, J. A., Belanger, J. L., Mathura, R. K., Meszéna, D., McGinley, M., Muñoz, W., Banks, G. P., Cash, S. S., Hsu, C.-W., Paulk, A. C., Provenza, N. R., Watrous, A. J., Williams, Z., Goldman, A. M., Krishnan, V., Maheshwari, A., Heilbronner, S. R., Kim, R., Rungratsameetaweemana, N., Hayden, B. Y. & Sheth, S. A.
• Publication: Nature
• Published: 06 May 2026
• DOI: 10.1038/s41586-026-10448-0
• URL: https://www.nature.com/articles/s41586-026-10448-0

🎯 HOOK

What if your brain keeps thinking, parsing sentences, and predicting what comes next even when you’re completely under? Baylor College of Medicine researchers inserted Neuropixels probes into the hippocampus of patients undergoing epilepsy surgery under general anaesthesia and found something that should not be possible according to most theories of consciousness: individual neurons were discriminating oddball tones, processing semantic categories of words, and even encoding information about upcoming sentences. The hippocampus was doing all this while the patients were fully unconscious.

💡 ONE-SENTENCE TAKEAWAY

The human hippocampus performs complex linguistic and predictive processing during anaesthetic-induced unconsciousness, demonstrating that semantic comprehension and representational plasticity do not require conscious awareness.

📖 SUMMARY

Background. A central question in cognitive neuroscience is how much of our higher-order processing depends on being conscious. Prominent theories of consciousness argue that sophisticated pattern recognition, semantic interpretation, and predictive processing require conscious access, especially when integrating across timescales. Anaesthesia provides a reversible, well-characterized state of unconsciousness that offers a clean test bed for this question.

What they did. The team recorded from 7 patients undergoing anterior temporal lobectomy for drug-resistant epilepsy. After the lateral temporal cortex was resected but before removing mesial temporal structures, they inserted Neuropixels 1.0-S probes (384 recording channels each) into the hippocampus. This was the first time Neuropixels had been used in this brain region; motion artefacts were markedly less conspicuous than in cortical recordings because the hippocampus is anchored by the dura of the middle fossa. They isolated 651 units across all patients.

The oddball experiment. Three patients heard sequences of pure tones (100ms, 200Hz vs 5kHz) with an 80-20 probability split, the rare tone was the oddball. The counterbalanced version swapped which tone was rare halfway through, with a 50-50 washout period in between. Results:

• 70.9% of units (122/172) showed tone-evoked responses. Neuronal responses differentiated standard from oddball tones.
• The divergence was sharpest within the first 300ms. 24.7% of units signalled oddballs.
• Local field potentials (LFPs) showed oddball-evoked responses: a negative deflection in ERP and increased gamma amplitude.
• SVM decoding of tone identity hit 0.61-0.70 accuracy across patients (P<0.001). Oddball identity was also decodable, though weaker.

Plasticity during unconsciousness. The oddball response grew stronger over the ~10-minute experiment. Splitting each block into halves showed a significant increase in oddball encoding. A 50-trial sliding window revealed continuous improvement correlating with trial position. The neural population vector for oddballs didn’t just gain amplitude, it rotated in high-dimensional space (cosine angle r=0.5). This is real plasticity, not just gain modulation.

An RNN model trained only to discriminate tone identity spontaneously developed oddball discrimination as an emergent property. Systematic lesioning showed that inhibitory connections (I-to-E and I-to-I) were essential.

The language experiment. Four patients listened to podcast episodes (The Moth Radio Hour, Kurzgesagt) while under anaesthesia. The researchers aligned neural activity to word onset/offset and found:

• Word frequency encoding. Neurons responded differently to rare vs common words. Significant positive correlation in single-unit activity across all four patients (mean r=0.48). This held even after controlling for word duration.
• Semantic embedding. A linear model predicting firing rates from semantic embedding vectors outperformed shuffled data in all units. Average correlation between predicted and actual firing rates was 0.397, higher than the 0.226 observed in awake patients from the same group’s prior work. Using only unique words, 75.4% of units still showed significant results.
• Semantic categories. 85.6% of units showed selectivity for at least one of 12 semantic categories. 63.7% discriminated 2+ categories. These proportions closely matched awake patients.
• Parts of speech. 79.5% of units distinguished nouns from non-nouns. Median was 4 out of 11 POS categories per neuron.
• Prediction. Neural responses correlated with the semantics of upcoming words. Future words were decoded nearly as well as past words. Surprisal, computed via GPT-2, modulated firing rates in 65.6% of units.

What it means. The hippocampus processes abstract semantic relationships, distinguishes grammatical categories, and anticipates upcoming linguistic content without any conscious awareness. The key ingredient for consciousness is not local neural activity in the hippocampus. Consciousness may involve cross-regional coordination (global workspace), global propagation of local signals, or recurrent processing. What this rules out is the idea that complex language processing itself requires consciousness.

💡 INSIGHTS

1. The hippocampus is doing heavy lifting outside awareness. It’s not just a memory structure that needs consciousness to function. It’s actively parsing, categorizing, and predicting sensory input even when you’re out cold.

2. Plasticity does not require wakefulness. The oddball discrimination improved over 10 minutes. The neural representation changed in shape, not just amplitude. This is the kind of learning we usually associate with being awake and paying attention.

3. Language prediction is a low-level property. The fact that hippocampal neurons encode upcoming word meaning and respond to surprisal suggests that prediction is baked into the architecture of the system, not something consciousness layers on top.

4. Anaesthetized brains look surprisingly like awake brains on language. The proportions of category-selective neurons, the semantic embedding regression performance, and the prediction effects all closely matched data from a separate cohort of awake patients.

5. Clinical implications. Post-anaesthesia implicit recall has been reported. This study offers a mechanism: sensory processing and plasticity are preserved even when explicit memory consolidation is knocked out by anaesthesia.

6. Not just low-level sensory. The oddball learning happened on a ~10-minute timescale, far longer than adaptation or repetition suppression. The semantic effects require processing abstract categorical structure, not just acoustic features.

🧠 FRAMEWORKS & MODELS

Layers of Processing Found

RNN Model

A continuous-rate RNN was trained on a signal-detection task with two input channels (tone A / tone B). Training had three stages: 80-20, washout 50-50, then 20-80. Despite training only on tone identity, the model spontaneously developed oddball discrimination, it emerged from the network dynamics.

Systematic lesioning revealed that inhibitory connections (I-to-E and I-to-I) caused the most severe drops in decoding accuracy. Inhibitory feedback is the critical mechanism.

📝 APPLICATIONS

For Neuroscience

• Challenges assumptions that consciousness is needed for semantic processing
• Suggests consciousness depends on cross-regional coordination, not local computation
• Mechanistic explanation for implicit memory under anaesthesia

For AI and BCIs

• Hippocampal prediction mirrors LLM behavior (GPT-2 surprisal matching neural responses)
• Speech prosthetics could potentially tap into hippocampal signals
• “Can we use these signals to deploy and run a speech prosthetic for parts of the brain damaged by stroke or injury?” - Dr. Vigi Katlowitz

For Clinical Practice

• Anaesthesia monitoring: the brain does more than assumed during surgery
• Post-operative implicit recall may be more common than recognized

⚠️ LIMITATIONS

1. Single anaesthetic. All patients were under propofol. May not generalize to other anaesthetics, sleep, or coma.
2. Small sample. 7 patients total. Hard to test for lateralization effects.
3. Single brain region. Only the hippocampus was recorded.
4. No behavioural readout. No way to test if this processing reaches conscious access.
5. RNN dependency. Model relies on binary input categorization inherited from other regions.
6. Prediction vs contextualization. Future-word effects could reflect contextualization rather than active prediction.

📚 REFERENCES

• Primary Source: Katlowitz et al. (2026). Nature.
• BCM News: Researchers discover advanced language processing in the unconscious human brain
• Prior work: Franch et al. A vectorial code for semantics in human hippocampus.
• Prior work: Katlowitz et al. Semantic contextualization in human hippocampus.

Related Reading:

• Good Sleep, Good Learning, Good Life by Dr. Piotr Wozniak - The sleep-learning connection
• Dehaene & Changeux (2011). Experimental and theoretical approaches to conscious processing. Neuron.
• Tononi et al. (2016). Integrated information theory. Nature Reviews Neuroscience.
• Mashour et al. (2020). Conscious processing and the global neuronal workspace. Neuron.

Crepi il lupo! 🐺