High-level Prediction of Continuous Speech During Mind-Wandering

Chen, Gal R.; Finkelstein, Rachel; Goldstein, Ariel; Hassin, Ran R.; Deouell, Leon Y. · 2025 · Crossref

DOI: 10.1101/2025.08.29.673052

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study investigates whether high-level speech processing, specifically context-based prediction and semantic encoding, persists during mind-wandering (MW), a state characterized by reduced subjective awareness of external stimuli. The research addresses a critical gap in understanding the relationship between predictive processing and consciousness, questioning whether the brain’s ability to predict upcoming linguistic content is sufficient for subjective experience or if it can operate independently when attention disengages. This inquiry is particularly relevant given current debates regarding the consciousness of artificial intelligence systems that utilize similar predictive computational mechanisms. The researchers recorded electroencephalography (EEG) data from 25 participants listening to 130 minutes of Hebrew audiobooks. Participants provided self-reports via random probes, labeling the preceding 12 seconds as either "on-task" (attentive) or mind-wandering. The analysis employed three primary metrics to assess neural processing: (1) spectral power changes to validate MW states; (2) neural encoding of word-level semantics using Word2Vec embeddings; and (3) predictive processing markers, including responses to word-level contextual surprise (quantified by Large Language Model probabilities) and the encoding of pre-word contextual embeddings (derived from the hidden layers of the Mistral LLM). Linear encoding models and time-resolved regression-based event-related potentials were used to compare neural activity between MW and on-task periods. The results confirmed that MW was associated with distinct spectral changes in theta and low beta bands and a significantly reduced early neural response to word onset, consistent with perceptual decoupling. However, high-level processing remained robust during MW. Semantic encoding of word meanings persisted without significant reduction compared to attentive periods. Crucially, the neural response to contextual surprise—indicating the brain’s reaction to unexpected words—remained intact and statistically indistinguishable from on-task responses, despite the attenuated onset response. Furthermore, the neural encoding of predictive contextual embeddings, which reflect the brain’s use of prior context to anticipate upcoming words, was strong during MW, though slightly weaker than during attentive listening. Control analyses ruled out acoustic features as the sole driver of these effects. These findings demonstrate that sophisticated, context-based language prediction and semantic processing continue during mind-wandering, even when subjective awareness of the speech content is diminished. This dissociation suggests that the computational mechanisms shared between human brains and Large Language Models are insufficient for generating the subjective experience of meaning. The study implies that predictive processing alone does not equate to consciousness, offering significant theoretical implications for predictive processing accounts of consciousness and debates regarding the potential for consciousness in artificial intelligence.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• mind wandering