Schönmann et al. — Stimulus Dependencies, Not Prediction, Explain Pre-Onset Brain Encoding
Summary: Antonello (eLife 2026) reviews Schönmann, Szewczyk, de Lange & Heilbron (eLife 2026): apparent neural encoding of upcoming words before they are heard is explained by language’s bidirectional statistical structure, not by genuine predictive computation in the brain. A methodological challenge to an entire field of brain-LLM comparison research.
Sources: Raw/Language Models_ Does the brain really know what word is coming next_.md
Last updated: 2026-05-07
The prior claim being challenged
Goldstein et al. (Nature Neuroscience 2022) found that word embeddings from a language model could predict brain activity before a participant heard the corresponding word. They and others interpreted this pre-onset encoding as evidence that the brain predictively computes upcoming words — a central prediction of predictive coding theory (Rao & Ballard 1999). The inference: brain and LLM share the same computational strategy.
What Schönmann et al. show
They reanalyzed two independent datasets using two explicitly non-predictive systems:
- A randomized lexical word embedding — captures only the semantic properties of the current word; cannot encode future words
- An auditory representation of the current stimulus — no word-level encoding at all
Both non-predictive systems produced the same apparent “predictive” signature as the brain showed in the original studies.
Reductio ad absurdum: if systems that provably cannot predict future words appear to do so under the original methodology, the methodology cannot serve as evidence for predictive computation in the brain.
Why this happens: bidirectional dependencies in language
Language is not a Markov chain. Adjacent words carry mutually constraining information in both directions:
- Collocations and idioms: “sour grapes” — knowing “sour” constrains “grapes” and vice versa
- Multiword expressions: “Statue of Liberty” — the first two words encode the third
- Syntactic cues: tense marking, passive/active voice spread information across clause boundaries
These bidirectional dependencies mean that a system encoding only the present word will implicitly carry information about neighboring words — including future ones. Prior methods attempting to correct for this (next-word residualization) fail to remove the artifact.
The methodological lesson
Computational encoding models are correlational, not causative. Finding that brain activity correlates with language model representations — even future ones — does not establish why that correlation exists. Mechanistic claims require causative experiments (e.g., lesion studies, TMS, carefully controlled naturalistic paradigms) that directly manipulate the variable of interest.
Antonello’s commentary situates Schönmann alongside a growing literature of caution (Antonello & Huth 2022; Guest & Martin 2023; Azizpour et al. 2025) questioning encoding-based inference in cognitive neuroscience.
Implications for the brain-LLM parallel
One prominent argument for taking LLM representations seriously as models of cognition is the empirical finding that LLMs and brains appear to share predictive processing strategies. Schönmann et al. undercut the empirical basis of that argument: the apparent sharing may be an artifact of language statistics, not evidence of shared computation.
This does not settle whether the brain uses predictive coding — it shows that the pre-onset encoding evidence cannot settle it. The question remains open; the methodology that was claimed to answer it doesn’t.
Connections
predictive-coding — the theoretical framework at stake; this paper is a significant challenge to the empirical evidence base for predictive coding in language processing.
llm-and-mind — the brain-LLM parallel is weakened by this finding. The claim that LLMs and brains share computational principles rests partly on studies Schönmann et al. now contest.
jaeger-relevance-realization — Jaeger et al. argue from a theoretical direction that the brain-as-algorithm equivalence is a category error. Schönmann et al. add an empirical direction to the same skepticism.
hawkins-a-thousand-brains — Hawkins posits the neocortex as a prediction machine using reference frames. Schönmann et al. question whether the standard experimental evidence actually demonstrates prediction at the word level.