NPF-02 · proposal · not executed
Talking to myself or to an AI
The same word can be a private thought or a message addressed to someone. We want to investigate that difference without confusing recognition with consent.
| Situation | Example | Expected output |
|---|---|---|
| Perception | Hearing or reading “up” | No transmission |
| Private speech | Saying “up” to myself | No transmission |
| Addressed speech | Saying “up” to the AI | Review and confirmation |
| Visual imagery | Imagining movement | No transmission |
The central question
Can we distinguish “I say up to myself” from “I say up to the AI”? The person may know their intention clearly, but this does not establish EEG detectability. Assigned tasks and self-reports are imperfect labels, not direct access to thought.
Thinking is not always inner speaking
Inner speech can accompany reading, writing and sentence preparation. We can also imagine scenes. Hearing or smelling something can prompt words or mental images: perception and imagination may overlap. We do not assume a unique EEG signal for each experience.
How we would test it
We propose comparing the same four words in private and assistant-addressed inner speech. Listening, silent reading, visual imagery and rest are controls; a naturalistic block includes writing and free thought. Olfaction would require a specific extension and would not be validated by these controls.
A manageable pilot
Proposal: 12 participants, two days, 20 trials per condition/day: 2,880 trials. Adding 30 negative minutes per visit gives 12 hours without communication intent. This is a feasibility estimate, not a power calculation or privacy validation. It requires ethics review, acquisition consent and suitable hardware.
Preventing shortcut learning
The model receives neither the activation button nor task markers. Cues are counterbalanced and excluded from analyzed windows. Outer evaluation holds out entire participants; transforms, hyperparameters and thresholds are fitted on training data only. EOG, EMG and timing cues receive their own controls.
What would count as progress
Proposed criteria, all required: mean balanced accuracy ≥0.60, lower 95% CI >0.50 and p<0.05; sensitivity ≥80% and one-sided 95% upper false-activation bound ≤1/hour. Per-person outcomes, abstentions and failures must be reported. These research criteria await preregistration; they are not safety guarantees.
Consent for every transmission
During research, activation is explicit and external. Text is reviewed and confirmed before leaving the device; errors, cancellation or timeout block transmission. No classification result substitutes for permission. The IS-01 pilot has not evaluated this distinction.


