James wears headphones that sense his brain’s electrical activity (EEG) while he works with Claude, Codex, and other AI models. He records that activity alongside his typing rhythm, dictation, and the timing of every prompt and reply, then publishes what he finds, including what didn’t work.
Could everyday EEG headphones ever tell an AI assistant that you noticed its mistake?
That’s the long question, and nothing here answers it yet. The groundwork comes first: a clean signal, honest timing, and a record of what changes together. This is one person exploring. Nothing here diagnoses, treats, improves focus, or reads thoughts, and one person’s results can’t speak for anyone else.
Sharing one clock lets changes in the EEG be lined up with what he was doing.
The study keeps these on his Mac and never publishes them. This page shows summary numbers and a few short test traces.
Every measurement on this page is read from the study’s small data files, updated as sessions are added. Normal work days are recorded during everyday work; focused research sessions are set up for the study.
The numbers load from the study’s data file. Turn on JavaScript to see them here. Open the data file
These come from the first session, a normal work day. Each finding carries a label for how much weight it can bear, with its numbers beside it.
How to read the numbers. Changes in EEG power are in decibels (dB), a log scale where zero means no change. The interval is the range the true change probably falls in; when it includes zero, the change may be nothing. A p value is how often shuffled data showed a difference at least as large. Windows are short stretches of EEG; any with too many filled-in samples, estimated by the headset’s software rather than measured, are set aside.
When James dictates, the sensors pick up his jaw and face muscles, and the muscle measure (fast activity that working muscles produce) jumps at each dictation start. Because those jumps land where the recorded dictation times say they should, the signal path and its timing work at the scale of seconds.
Keep in mind: This is muscle, not brain activity, so dictation is analyzed on its own.
Numbers appear here with JavaScript on.
Alpha, a rhythm that usually rises when the eyes close, ran a little higher while James waited for a reply than while he read it. Reading tends to lower alpha, so the direction makes sense, and the difference held with speaking set aside.
Keep in mind: It’s one result among many comparisons, and eye movement differed too, so it needs its own test.
Numbers appear here with JavaScript on.
Over the current Bluetooth path, the EEG arrives late and in bursts after connection hiccups, and the delay wanders during a session. The headset numbers its samples and none were skipped, so one steady rate fits the whole session, but each sample can only be placed in time roughly. Quick brain responses unfold within fractions of a second, so the goal is timing within 20 milliseconds.
Keep in mind: This measures when samples arrive, not the full delay from head to file, which isn’t measured yet.
Numbers appear here with JavaScript on.
During short bursts, the headset’s software filled in samples, meaning it estimated them instead of measuring them. The rest of the session had almost none. Typing, clicking, sound, and the Mac’s workload didn’t explain the bursts, and one stopped soon after James switched off his iPhone’s Bluetooth.
Keep in mind: That’s one example, not proof.
Numbers appear here with JavaScript on.
A basic EEG check: close the eyes, and alpha should clearly rise. It rose only a little, and its interval included zero. The headset also rated some sensors poor for the whole session.
Keep in mind: Until this check passes, treat every EEG finding here as provisional.
Numbers appear here with JavaScript on.
Each of these comparisons had an interval that included zero, and typing speed didn’t follow any EEG rhythm.
Keep in mind: With timing this loose, quick brain responses would be smeared out, so this says little about them.
Numbers appear here with JavaScript on.
Tokens turn each short stretch of EEG into one symbol from a small, fixed alphabet of patterns, much as text is split into tokens for AI models. Most stretches were set aside because they held filled-in samples, so less data loss is the main way to get more. Token use looked slightly different across phases of work.
Keep in mind: The samples are small, so those differences are only a hint.
Numbers appear here with JavaScript on.
Bars show intervals, ticks show estimates, and the dashed line marks no change. Drawn only from the data file.
This time an eye-tracking camera joined the EEG headphones and the key and mouse logger, all on one Mac clock. A series of short, structured tests checked what each one picks up and how well the three line up.
James wore the EEG headphones, with an eye-tracking camera between his two screens, while a logger recorded key and mouse timing, never the characters typed. This page shows numbers estimated from the camera, not its images.
Numbers appear here with JavaScript on.
Colored parts are the test blocks, in order.
Between blocks came setup, an app restart, and normal work in other chats. That makes the whole recording a normal work day in the sessions table, so this section uses only the test blocks.
These numbers come from the test’s own data file. Open the test data
How to read this test. It’s one person, it’s exploratory, and it isn’t blinded, since James knew what each test was for. These are signal checks, not claims about people in general, diagnosis, or mind reading. EEG sizes are in microvolts (µV), as recorded and unverified, and gaze positions are unvalidated estimates. A t value weighs a difference against its scatter, so a larger one is less likely to be chance. Match strength is the size of a correlation, r, from zero for no match to one for a perfect match; a negative r means two traces mirror each other. Screen points are the units macOS uses to lay out a screen.
I wore my EEG headphones and an eye-tracking camera at the same time, with my keystroke timing logged on the same clock. Here’s what thirty-one minutes taught me.
The headphones can see where my eyes point. When I looked left, then right, the sensors on my right ear swung about a hundred and thirty microvolts, a mirror image of what the camera saw.
The two clocks agree. The EEG and the camera put each eye movement at the same moment, within about five milliseconds.
Blinks don’t show up. The camera caught every one, but the headset measures against my left ear, and a blink probably reaches both ears about equally.
The eye tracker picked the right screen eighty-three percent of the time. Good enough for big areas, not single words.
And it was my cleanest recording yet: less than one percent of the brain data had to be filled in, down from about seven percent on October twentieth.
What’s next is the moment that matters most: noticing when an AI gets something wrong. Headphones that can sense that could let an assistant correct course right away. One person, one session. It’s a start.
When a spoken cue asked James to look far left or far right, with his head held still, the right-cup sensors swung one way for left and the other way for right. Every right-cup sensor moved, and the left-cup sensors moved far less. The headset measures every sensor against a reference point on the left cup, so the right-cup sensors span the head from ear to ear.
Keep in mind: Eye and muscle activity dominate EEG measured around the ears, so this is an eye-movement signal picked up by the headphones, not a brain signal.
Numbers appear here with JavaScript on.
The right-cup EEG and the camera’s reading of the pupil, on one clock. Labels along the top mark each spoken cue to look left or right. The two traces mirror each other.
Matching how fast the eyes moved, as seen in the EEG and by the camera, gives the best fit with no shift between the two. Leaving out any one look barely moves that best match. The eye recorder and the EEG recorder also share one clock.
Keep in mind: Both paths still carry their own delay from the real eye movement, so this shows they match each other, not that either delay is zero.
Numbers appear here with JavaScript on.
Match strength between eye-movement speed in the EEG and in the camera, with one recording shifted earlier or later against the other. The dot marks the best match.
The camera saw the cued blinks, but the right-cup EEG, averaged over the blinks, stayed inside the range it shows at random moments. The likely reason: a blink reaches both ears about equally, so measuring against a reference on the left ear cancels it out.
Numbers appear here with JavaScript on.
How open the eyes were, as a share of normal, and the right-cup EEG around each cued blink. Averaging the blinks makes the dip look shallower than a single blink.
After a fresh calibration, James looked at test dots the tracker hadn’t been calibrated on. Its estimates usually landed on the correct screen, but a typical miss covers a large patch of screen. An earlier check with the same app missed by more but picked the correct screen more often.
Numbers appear here with JavaScript on.
Each ring is a test dot, joined to the tracker’s typical estimate for it. Faint dots are single estimates, which count as correct only inside their test dot’s screen.
Alpha, the rhythm that usually rises when the eyes close, rose with eyes closed in the rest baselines at both the start and the end. Across the session, the rise has an interval that stays above zero, a first for this study; in the first session it included zero. The rise was strongest on one left-cup sensor.
Numbers appear here with JavaScript on.
Fewer samples were filled in by the headset’s software than in any earlier recording, and sample arrival times spread far less around one steady rate.
Keep in mind: The short pilot between the two sessions was saved in a different format, so its timing can’t be compared.
Numbers appear here with JavaScript on.
No mistake was flagged during the reading test, so there are no “noticed a mistake” moments to study yet. Small, everyday eye movements, while reading and during calibration, didn’t track the EEG beyond chance. The tracker also had gaze estimates for only part of the reading.
Numbers appear here with JavaScript on.
Timing shows how widely sample arrival times spread around one steady rate. The eyes-closed check is the basic signal test described above.
| Date | Kind | Minutes of EEG | Filled in | Timing (±) | Eyes-closed check | Dictations |
|---|---|---|---|---|---|---|
| Session numbers appear here with JavaScript on. | ||||||
The aim is still the question at the top of this page. The combined test shows the hardware and clocks are good enough to line up eye movements, EEG, and typing on one timeline, and that eye movements must be modeled as their own signal. Here’s what changed after the first session, and what comes next.
Run the mistake-spotting reading test again, with a clearer routine for pressing the mistake hotkey.
Detect each eye jump during reading, so eye movements can be modeled as their own signal.
Repeat sessions to see whether these results hold.
Record the first focused research session, set up around the first session’s problems: phone Bluetooth off, the headset app’s own file recording off, the poorly rated sensors pressed flat, and a short reading-only block to test the waiting-versus-reading lead.
Check across more sessions whether keeping the phone’s Bluetooth off prevents the bursts of filled-in samples.
Automatically mark moments when something felt off, such as a stopped reply, a quick correction, or a prompt that opens with “no” or “actually”, then compare the EEG around them.
Look for tighter timing: ask the headset’s maker how filled-in samples are marked and estimated, what its sample rate means, and whether each sample can carry its own timestamp.
Spot a flaw in the method, or have an idea for the next session? Tell James.
Interested in the Mac-local loop, the iPhone companion, early testing, or the research behind it? Start a conversation with James.
james@noticingmind.com · The email link opens your mail app.