Consent and Privacy for First-Person Video Datasets: A Practical Checklist

The short answer: A first-person video dataset is only usable if its consent and privacy process is sound. Contributors should receive a plain-language explanation of what will be recorded, how it will be used, who will see it, how long it will be kept and how to withdraw; consent should be recorded with each recording; protocols should keep bystanders, screens and documents out of frame; every session should be reviewed for privacy issues; and withdrawals should remove data everywhere it went. Teams often run this through managed egocentric data collection programs.

Why consent matters more for first-person video

Egocentric video captures more than a task. A camera on someone’s head records their hands, voice, home, workplace, family members, screens, mail and anything else in view. It is personal data under GDPR and similar laws, and in some uses it may involve biometric data, which carries extra obligations. A dataset collected without clear consent may be unusable, however good the footage.

Partners and customers increasingly ask how training data was collected. A documented consent process, applied consistently, is what lets a team answer with confidence and keep using the data as rules evolve.

Explain the data use in plain language

Consent is only meaningful if contributors understand it. The explanation should say what the camera records, what tasks they will perform, how recordings will be used, such as training AI models, who will access the data, whether it may be shared with partners, how long it will be kept and how to withdraw. Avoid legal language where a simple sentence will do.

Offer the explanation in the contributor’s language and give time to ask questions. Recording that the explanation was given, and in which version, is part of the consent record.

Record consent with every recording

Consent should be documented and linked to each recording through its metadata. That means storing the signed form or digital consent, the version of the explanation, the date and the scope agreed, such as which tasks, settings and uses are covered. When a recording moves into a dataset, its consent record should travel with it.

This link makes audits and withdrawals possible. Without it, nobody can say which recordings a given consent covers. See AI data collection for consent practices across other data types.

Keep bystanders out of frame

People who did not consent should not appear in recordings. Protocols can reduce the risk: record when others are not present, choose settings where bystanders are unlikely, ask household members to consent or leave the room, and avoid public spaces unless the project specifically requires them. Contributors should be briefed with examples of what to avoid.

Where children may appear, stricter rules apply, and many programs exclude them entirely. When bystanders do appear, review should blur, crop or reject the affected segments according to the rules agreed with the privacy team.

Screens, documents and identifying details

First-person cameras easily capture phone notifications, computer screens, letters, bank cards, photographs on walls, street names and house numbers. Protocols should tell contributors to turn off notifications, put documents away and avoid filming identifying details. Some programs provide a clean workspace or set for recording.

Even with good protocols, details slip through. Review checklists should cover screens, documents, faces in photographs and location identifiers, and reviewers should blur or reject as required.

Review every session

Automated face and text detection helps, but human review catches what tools miss: a reflection in a window, a name on a parcel, a voice of someone off-camera. Reviewers should check every session against a written checklist, apply agreed treatments consistently and log what they changed. Rejection reasons feed back into contributor briefings.

Quality review can run at the same time, checking whether the task was performed as scripted and whether the footage is technically usable. See data labeling for how reviewed footage moves into annotation.

Make withdrawal real

Contributors should be able to withdraw consent, and withdrawal should actually remove their data. That requires knowing every place a recording went: raw storage, processed datasets, annotation tools, partner copies and training sets. A tracking system that records where each recording is used makes removal practical.

Explain at consent time what withdrawal can and cannot undo, for example whether models already trained on the data will be retrained. Clear, honest answers build trust and reduce disputes.

Fair payment and clear terms for contributors

Contributors give time, effort and personal information, and they should be paid fairly and on time under clear terms. The terms should explain how payment is calculated, what happens if a session is rejected for quality reasons, how devices are handled and returned, and how to raise concerns. Payment should never depend on contributors filming more than the protocol allows or including other people.

Clear terms reduce disputes and improve data quality, because contributors who understand the rules follow them. Payment and labor rules differ by country, so programs collecting in several markets need local checks.

Sharing datasets with partners and customers

Many egocentric datasets are licensed to partners or used across several teams. Consent should cover these uses explicitly, and sharing agreements should require partners to respect the same retention, privacy and withdrawal rules. Datasets shared outside the original team should carry documentation describing how they were collected and what consent covers.

When a withdrawal request arrives, partners holding copies need to be notified and asked to delete the affected recordings. Contracts should make this obligation clear before any data changes hands.

Running consent as part of operations

Consent and privacy work best when they are built into collection operations rather than handled by a separate team at the end. Recruitment, onboarding, device setup, recording, review and delivery each have a privacy step, and each step should be checked in quality reviews.

Our egocentric data collection outsourcing page explains how vetted Philippine teams run contributor programs with consent and privacy review, and why diversity beats volume in robot demonstration data covers a related data type in the same cluster.