Ethics & consent.
This competition follows the NeurIPS Code of Ethics, with explicit consent, controller, privacy, and read-only decoding commitments.
Following the NeurIPS Code of Ethics on consent, privacy, and representativeness.
Non-invasive neural decoding has clinical value, but raises a mental-privacy concern when applied to non-consenting subjects.
Quoted from the competition proposal:
The task design also follows the NeurIPS Code of Ethics.
- Track 1, EEG-to-IMG. Decoding controlled visual stimuli the participant is actively viewing. Does not test imagination, memory, or private thought.
- Track 2, BCI decoding. Decoding cued mental tasks (motor imagery, mental calculation, word association). Each command is a deliberately chosen action, not unconstrained inference.
- Track 3, Sleep onset. Decoding from at-home wearable EEG with subject-initiated recording. The subject controls when the headband is worn.
- Track 4, EMG-to-Pose. Regressing hand-joint trajectories from wristband EMG while participants deliberately perform instructed movements recorded with motion capture.
Every track operationalises consent through controlled-stimulus, cued-task, participant-directed movement, or home-wearable setups, never covert observation.
Why non-invasive decoding is privacy-bounded today.
Two empirical limits bound mental-privacy risk in non-invasive recordings. First, decoding accuracy is high only for perceptual tasks - viewing an image, executing a cued movement - and falls sharply on imagined content. Second, when the subject diverts attention from the cued task (counting backward instead of attending the stimulus), the recoverable signal collapses and accuracy returns to chance.
Consent is therefore a technical precondition, not only a legal one: a non-consenting subject defeats decoding by attending elsewhere. Each track enforces consent through its setup - controlled-stimulus, cued-task, participant-directed movement, or home-wearable - and never through covert observation. Decoders in this competition are read-only, with no closed-loop actuation, and all evaluation is open and peer-reviewed.
Accuracy collapses without subject engagement. Compliance is a precondition, not an afterthought.
Every dataset has an identified controller and a consent chain.
Every human-subject dataset comes from its original data controller. We use it only after confirming that recordings, annotations, and competition use satisfy the relevant consent, ethics approval, de-identification, and data-use terms, alongside the NeurIPS Code of Ethics.
- Track 1, EEG-to-IMG. THINGS-EEG1/2 (THINGS consortium), Alljoined-1, Alljoined-1.6M (Alljoined).
- Track 2, BCI decoding. Stieger 2021, Dreyer 2023, Zyma 2019, Scherer 2015 (original lab data controllers); 20-subject Graz/BrainHero release (Inria Bordeaux, NEARBY/PROTEUS projects).
- Track 3, Sleep onset. Sleep-EDF Extended (PhysioNet), PhysioNet Challenge 2018 (PhysioNet), HMC Sleep Staging (Haaglanden Medisch Centrum), Muse Sleep-Onset (Muse).
- Track 4, EMG-to-Pose. Salter2024 EMG2Pose (Meta Reality Labs); held-out participant and movement-stage evaluation splits.
Each controller approves the recording paradigm, the annotation protocol, and the competition use separately.
Tracks do not launch until provider approvals are finalised.
For unreleased hidden data, the corresponding track does not launch until data-use permission, anonymisation, and approval for competition use are finalised. No personally identifying information is exposed to participants at any phase. The datasets are not intended to be demographically representative of all users or clinical populations; claims are limited to the generalisation axes explicitly tested by each track.
Results extrapolate only along the generalisation axes each track explicitly tests, not to populations or settings outside.
What this competition will not do.
Three commitments are non-negotiable. They constrain both the design of the competition and the artefacts released after it.
-
Closed-loop NO
No closed-loop actuation.
All decoders in this competition are read-only. No track involves any actuator, BCI command output to a physical device, or feedback loop into the subject.
-
Non-consenting subjects NO
No decoding from non-consenting subjects.
Every participant in every dataset has provided explicit consent. The competition does not accept submissions trained on covertly collected data.
-
Dataset resale NO
No resale of any released dataset.
Released data carries the original provider's licence. Re-use is permitted under that licence; resale and re-licensing are not.
Ethics or consent questions?
Email the organising committee or open a thread on Discord.