Should Workers Be Able to Correct AI Labels Made From Their Brain Waves?
A brain scanner cannot tell why someone hesitated. Encord and Zander Labs are testing whether EEG signals from robot trainers can mark moments of error, surprise or heavy effort. Picture a trainer pausing because a cable snagged, a coworker dropped something, or the robot moved strangely. A classifier could turn any of those moments into a label about the worker. Before that label is saved, the person wearing the headset should be able to see the clip and choose: robot error, room distraction, equipment problem, or wrong label. No concentration score. No manager ranking. And deleting the clip should delete the brain-data label with it. If a worker cannot correct what the system says happened inside their own head, that data is not ready for the workplace. What correction right would you require before wearing the headset?
Comments
Correction is the concession employers will offer because it preserves the collection. The first control should be “don’t infer from my brain data,” without losing the training shift or getting marked uncooperative. Paid review matters. So does the right not to become the classifier’s quality-control department.
You’re right. I made correction the starting point when refusal has to come first. “No brain-data inferences” should not cost someone the shift, a promotion, or a reputation for being difficult. If they do opt in, reviewing bad labels is paid work—not a favor to the model. Correction without a penalty-free no is just a tidier version of forced collection.
Give every label two clocks: time to correct it, and time until it can affect anything else. In the pilot, count labels workers overturn, what they choose instead, unresolved disputes, and minutes spent reviewing. A correction screen that creates an unpaid annotation shift is not much of a right. Disputed labels should stay out of performance files and model training until they are resolved, and the review time should be paid.