Training data licensing and consent: a buyer's checklist

Why provenance matters

Provenance is the record of where data came from, who made it and on what terms. You need it because a licence label on a download page is only a claim. The Data Provenance Initiative audited widely used AI datasets and found licences often miscategorised on dataset hosting sites, with licence omission of 70% or more and error rates of 50% or more. Those figures apply to the datasets the authors audited and say nothing direct about video datasets. The lesson carries over anyway: read the upstream terms yourself.

With custom collection you ask the same question earlier, before anything is recorded. Our egocentric video data page shows what a consented, reviewed custom collection looks like. The checklist below works for any supplier.

Licence types you will meet

Being able to download a dataset doesn't give you commercial permission. And a dataset that permits product development may still forbid you from passing the data on to anyone else. Our comparison of the main public egocentric datasets notes the terms of each.

Consent and releases

A licence sets your rights against the dataset's owner. Consent concerns the people in the footage, and with egocentric video that often means more than the wearer, because coworkers, customers and family members walk through the frame. For the recorder's side of this, see our guide on privacy and consent when recording at work, written for the people who record.

You can only check consent if the documentation lets you. Datasheets for Datasets asks every dataset to document its collection process, and Data Cards record upstream sources and collection methods. For wider governance, the NIST AI Risk Management Framework is voluntary guidance you can use to frame these questions. It is neither a licence nor a certification. Privacy law where the footage was recorded may also apply, and that's a question for your counsel.

Questions to ask any supplier

Ask these in writing, and keep the answers with the delivery:

1. Who recorded the footage, where, and under what agreement?
2. Did every identifiable person sign a release? How are bystanders handled?
3. Did every commercial site give written permission?
4. How are people under 18 handled? Are they excluded unless a parent or guardian signed?
5. What uses does the licence or contract permit? Is sublicensing allowed?
6. Are you barred from identifying anyone in the footage? Is any biometric processing done?
7. What personal data comes with the video: names, contact details, location?
8. What documentation comes with it? Look for a datasheet, a data card or something equivalent.

Put the answers next to your technical requirements: our egocentric dataset spec checklist covers the format side, and our checklist for judging a video dataset covers quality.

Common mistakes

• Trusting the licence label on a hosting page. Check the upstream terms, since the audit above found frequent errors.
• Reading "free to download" as "free to use commercially". Some research datasets, such as EPIC-KITCHENS, are licensed for non-commercial use only.
• Signing up to a draft. Use the executed agreement that names your organisation.
• Forgetting bystanders. The wearer's consent doesn't cover the people around them.
• Mixing sources without tracking their terms. If you pre-train on public video and fine-tune on robot data, as in egocentric video compared with teleoperation data, keep each source's licence attached to it.
• Choosing custom collection only for the licence. A commercial licence for public data is sometimes enough; see when public data is enough.

How TaskCapture handles consent

Collectors must get a signed release from every identifiable person they record and, in any commercial setting, written permission from the site's owner or manager. A person under 18 appears only with a release signed by a parent or guardian; otherwise the footage is rejected and deleted. TaskCapture does not run face or voice recognition and does not create biometric identifiers, and datasets never include collectors' names, contact details or payment data. Collectors agree to each project before they record, and footage from data suppliers follows the same consent rules. Buyers' contracts limit use to developing, training, evaluating and improving AI and machine-learning models, and bar them from identifying anyone who appears. The full rules are in our Privacy Policy.

Questions