Thought-to-Text

Thought-to-text is the consumer-facing term for the research direction of producing written or spoken language from neural recordings without going through the peripheral motor system — the model receives a multi-channel brain signal (EEG, fNIRS, MEG, or intracortical array) and emits text directly, bypassing keyboard, mouse, voice, eye tracking, or any other overt behaviour. The phrase was revived in mainstream AI coverage in Aug 2026 by Naomi Bashkansky, a former OpenAI founder, who positioned the goal as the telepathy endpoint: read neural activity, output the intended sentence. The defining operational claim is that the same scaling playbook that took text-from-text and text-from-speech from research demos to production models — large data, large neural network, no hand-crafted feature pipeline — should work for text-from-brain once the dataset is large enough.

The mechanism is the bitter lesson applied to a new modality. Today's brain-to-text decoders fall into two camps. The first uses intracortical electrodes (Utah arrays in motor cortex, recent high-channel-count intracortical microelectrode arrays) and reaches high accuracy on small vocabularies — the parallel arm of the field includes the BrainGate / Synchron / Neuralink demonstrations of typing-by-imagined-hand-movement, where the user thinks about moving a cursor and the decoder translates the motor-cortex signal into a keystroke. Thought-to-text is the more ambitious sibling: the user thinks of words and the decoder skips the cursor entirely, going thought → text rather than thought → motor → text. The second camp uses non-invasive signals (EEG, fNIRS, MEG) and reaches much lower accuracy but at zero surgical risk. Bashkansky's framing subsumes both: the claim is not a specific decoder architecture, but that scale across modalities — hundreds of thousands of hours of paired brain-and-text data, with model sizes comparable to frontier language models — should close the gap to production accuracy.

The discipline has prior art the entry should sit near: eeg-to-text is the academic standard for the non-invasive case, including the 2022-2024 wave of EEG-based sentence decoders (MetaAI brain2text, CMU corticomotor decoders, the Brain2Qwerty foundation-model work). bci-brain-computer-interface is the umbrella. wbe-whole-brain-emulation is the longest-horizon cousin. cuespeak benchmark and timespeak are BCI-speech benchmark cousins. thought-token is an adjacent 2024 concept describing a learnable representation of intermediate reasoning in language models — same word but unrelated. thinking-tokens is similarly a model-internal concept, not a brain-decoding one. The read/write symmetry worth flagging: thought-to-text without the inverse — text-to-thought — leaves a one-sided system, and the regulatory and consent calculus fundamentally changes once the model can in principle reconstruct what the user is hearing or reading in addition to what they are producing.

Tradeoffs and open questions. The bitter-lesson framing is plausible but unproven in this modality — the academic eeg-to-text work has not yet shown the same data-scale curve the bitter lesson predicts, because the data does not yet exist at scale and the relevant bottleneck is brain-recording hardware, not algorithms. Whether datasets of the size and quality needed can be collected ethically — the standard blocker for any neural-recording study — is unresolved. The technical claim that scaled compute and data will produce the result is conditional on several non-ML factors (electrode density, recording stability over time, dataset licenses) that have nothing to do with the standard LLM scaling narrative. The use-case motivation — restoring speech/locomotion to patients with ALS, locked-in syndrome, brainstem stroke — has long been the field's traditional justification, but Bashkansky's telepathy framing implicitly extends the use case to able-bodied users, which raises a different consent and cognitive-privacy register. From the AI-canon perspective, the entry is interesting primarily as a modality extension — if the scaling playbook does generalise to brain signals, it implies the same patterns from text/audio (RLHF on neural feedback, agentic tool use over neural input, multi-model ensembling) become applicable, and the public-system implications become substantially heavier than the assistive-medicine framing alone.