Dataset terms of use

ConsentVid datasets have no blanket licence. Each clip may be used only within the terms of its own consent receipt: a record in the contributor's ATproto repository, signed by their key, referenced from the manifest as cv:consentReceipt.

What a receipt grants

purposes: what it may be used for, e.g. training, evaluation, benchmarking. (researchOnly is deprecated; older receipts that use it keep their literal meaning.)
modelClasses: who may use it, e.g. academic, openWeights, commercial. (foundationModel is deprecated; older receipts that use it keep their literal meaning.)
modalities: which signals, e.g. video, audio, transcript, face, voice.
expiresAt: optional end date of the permission.

A use is permitted only if the receipt lists every applicable purpose, model class and modality. Anything not listed is not granted.

compensation records what the contributor said they expect. It is a declaration, not a payment arrangement: ConsentVid neither brokers nor enforces payment.

People in the video and age

Every receipt states who identifiably appears (peoplePresent: only the contributor, everyone shown having agreed, or no identifiable people) and that the contributor is an adult able to consent (adultAttested). These are self-declared under the contributor's key. A receipt without them is invalid. If you find a clip that contradicts its declaration, for example identifiable bystanders in a clip marked onlyMe, stop using it and report it.

Bots, AI-usage preferences and labels

A receipt from an account that labels itself a bot (the bot self-label on its profile) is not valid consent. Nor is a receipt granting training from an account whose own published AI-usage preference (community.lexicon.preference.ai) denies training: the contributor's preference wins. Clips the contributor labeled porn, sexual, nudity or graphic-media are left out of datasets unless you explicitly include them. resync checks all of this and drops such clips.

Withdrawal

Contributors can withdraw consent at any time by deleting the receipt (consent to processing personal data must be withdrawable, GDPR Art. 7(3)). Before every use, and at least before each training run, run:

consentvid resync manifest.json --write manifest.clean.json

and use only the clips that remain. resync verifies each receipt against the contributor's signed repository; withdrawn, expired or modified clips must be deleted from your copies. Clips that resync could not check are dropped too.

For contributors: records are public

Your video and consent records are public ATproto records. Relays, Jetstream and anyone indexing the network copy them as you write them. Deleting a record withdraws consent, and well-behaved indexes drop it, but deletion is not erasure: copies made before it can persist. Only the metadata you publish is in the records (title, description, tags, hashes, the storage pointer and the consent terms), so publish nothing there you would not want copied. Deleting or deactivating your ATproto account does not delete the video bytes on R2, IPFS or PeerTube: run consentvid revoke <video-uri> --delete-bytes for each video first.

Models trained before a withdrawal

Withdrawal does not affect the lawfulness of use before it (GDPR Art. 7(3)). A model trained in good faith on clips that were valid at a resync made before training stays legitimate and need not be destroyed. From the moment of withdrawal, the clip may not be used again: not for new training, retraining, fine-tuning, continued training from a checkpoint, or new evaluations. Delete it from your copies.

Attribution and authorship

Cite the dataset as given in the manifest's citeAs. Contributors attest, under their signing key, that they recorded each clip or otherwise hold the rights to grant consent. If you believe a clip was contributed without that right, report it to the contributor's PDS operator and stop using it.