Skip to content

Privacy and preparation

Robotics video worker consent workflow: notice, opt-in and exclusion

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A consent workflow for workplace demonstration video tells workers what will be filmed, why, who receives it and how to leave before any camera runs; records each person's yes or no under a pseudonymous ID; keeps people who decline physically out of the capture zone; reviews every clip for anyone without a recorded yes; and traces a later withdrawal through every clip and derived file.

Video and sensors

Robotics video worker

This page covers robotics video worker consent workflow: notice, opt-in and exclusion.Illustrative

Key takeaways

  • Tell workers in a fixed order: scope first, then a group briefing, then a written notice, then a private yes or no, and only then the cameras.
  • In an employment setting a yes only means something if saying no visibly costs nothing, so the decline path must be as easy as the opt-in.
  • Keep people who decline out of the capture zone by scheduling and roster, not by marking them; visible markers single people out.
  • Every clip is reviewed against the opt-in register before it leaves the site, and anyone without a recorded yes is cut, redacted or the clip is discarded.
  • A withdrawal must be traceable by pseudonymous operator ID to every clip, pose file and label, with the cut-off for data already delivered set in the agreement.

It needs five things in a fixed order: a written scope of what will be filmed, a briefing that reaches every worker who could appear, a private and recorded yes or no from each person, capture-day rules that keep people who said no out of frame, and a withdrawal route that still works after the footage has left the building. Miss any one and the archive carries people who never agreed.

Robot-learning footage makes this harder than ordinary workplace video because the person is the content. A demonstration clip is valuable precisely because it shows hands, posture and timing, so the training value and the identifiable person sit in the same frames. Blurring a face later does not fix footage of someone who was never asked.

This page is the operating procedure a site runs. It assumes the legal basis has already been settled with counsel; the pages on recording workers for AI datasets cover which notice, consent and biometric rules apply. Here the question is practical: who is told what, in what order, how the answer is recorded and how a no or a later change of mind is honored on the floor and in the files.

Why an employee's yes only counts if no costs nothing

Because the employer controls pay, shifts and promotion, a worker's agreement is only credible when refusing has no consequence at all. The European Data Protection Board says the imbalance of power in employment means employees can give free consent only in exceptional circumstances, where giving or refusing it has no adverse effect, and its video guidance adds that employers should generally not rely on consent for workplace video.

That does not stop a business from filming demonstrations. It means the program must be designed so that declining is ordinary: the same pay whether or not someone is filmed, no supervisor asking why, no list of decliners visible to the floor, and a non-filmed task available during capture shifts. Counsel decides which legal basis applies; operations makes sure the choice is real either way.

In the U.S., rules depend on the state and on what is captured. New York's electronic monitoring statute, for example, requires written notice at hiring, an acknowledgment and a posted notice for monitoring of telephone, email and internet use, which shows how narrowly some notice laws are drawn. If hand or face geometry may later be extracted from the video, Illinois's Biometric Information Privacy Act requires written notice of the specific purpose and length of term and a written release before a biometric identifier is collected. Whether pose data counts is a question for counsel, which is why the workflow asks it before filming rather than after.

The worker communication and exclusion workflow

The table is the asset: ten stages, each with an owner, the record it leaves and the point where a person can leave the program. Run it in this order; skipping ahead to cameras is the most common way a site ends up with footage it cannot license.

The worker communication and exclusion workflow
StageWhat happensOwnerRecord keptWhere a person can leave
1. ScopeWrite down tasks, areas, shifts, streams (video, depth, audio), derived data such as hand pose, recipients and retentionProgram lead with counselSigned scope sheet, version datedNot yet applicable
2. Briefing and noticeExplain the scope in plain language at each shift, including agency and night staff, then hand out a written notice in every language used on site; allow several working days before sign-up closesProgram lead, with HR or privacy for the noticeBriefing dates and attendee counts; notice text, version and distribution logAnyone can ignore the sign-up with no follow-up
3. Private decisionEach worker answers yes or no on a form returned to HR or privacy, never to the shift supervisorHR or privacy leadOpt-in register keyed by pseudonymous operator IDAnswer no, or do not return the form
4. Agency and contractor checkConfirm the agency or contractor allows its staff to take part and that its own notice reached themProcurementAgency confirmation letterTreated as no until confirmed
5. Capture rosterBuild shifts and zones so only opted-in workers are in the filmed area during capture windowsOperations schedulerRoster per capture windowAsk to be moved to a non-filmed task
6. Capture daySigns at zone entrances, a visible camera-on light, a stop word anyone can useCapture leadSession log with start, stop and pause timesSay the stop word or step out; recording pauses
7. Clip reviewCheck every clip against the roster and register; cut, redact or discard anyone without a recorded yesReviewer outside the line teamReview log per clip with action takenNot applicable; review is automatic
8. Derived dataGenerate pose, hand or gaze files only for operators whose yes covers derived dataData engineerDerived file manifest by operator IDDecline derived data separately on the form
9. WithdrawalA request by form or email removes the operator's clips and derived files from site storage and any undelivered batchPrivacy leadWithdrawal log with date and files removedAny time, by the same route used to opt in
10. After deliveryPass the withdrawal to the licensee under the agreement's deletion and model termsPrivacy lead with counselNotice sent and acknowledgment receivedAny time; the agreement sets what can still be removed

What workers hear, and in what order

Workers should hear the full scope in a group setting first and decide privately later, so nobody is asked on the spot in front of a supervisor. The UK Information Commissioner's Office says workers must be told about monitoring before it begins, in an accessible form, and that a data protection impact assessment can include discussing the plans with workers; the same order works well outside the UK.

Steps

What workers hear, and in what order

The written notice should be specific enough that a worker could explain it to a colleague. Vague purposes such as "recordings may be used for training" read as staff training to most people, so name AI and robot model development and outside recipients directly.

  • What is filmed: the named tasks and areas, the camera positions (fixed, head-mounted or wrist-mounted) and whether audio or depth is captured.
  • What is made from it: hand pose, body pose, gaze or object labels, and whether any of these could count as biometric data where the site is.
  • Who receives it: the business, its processing contractors and outside AI developers under a license.
  • How long it is kept, and what happens to it if the program ends.
  • That declining or withdrawing has no effect on pay, hours, assignments or reviews, and who to contact if anyone suggests otherwise.
  • How to withdraw later, by the same form or email used to opt in, and what can and cannot be removed once footage has been delivered.
  • Whether faces will be blurred, and the honest limit: hands, tattoos, uniforms and gait can still identify someone to colleagues.

Keeping people who said no out of frame

The reliable way to keep decliners out is to schedule them out, not to mark them. Colored lanyards or stickers for people who said no single them out in front of colleagues and undercut the point of a free choice. Build the capture roster so that only opted-in workers are assigned to the filmed zone during capture windows, and put signs at every entrance so anyone else knows the zone is live before they walk in, the same first-layer approach the EDPB recommends for video signage.

Even with a clean roster, someone will cross the frame. Decide in advance what happens to each kind of person, and log the action per clip.

Keeping people who said no out of frame
Who appearsDefault actionFallback if the default failsWhy
Opted-in worker, within scopeKeepNot applicableRecorded yes covers this task and stream
Worker who declined, crossing brieflyCut the segmentRedact face and body, then re-reviewNo recorded yes; their choice has to hold in the files
Worker who declined, central to the clipDiscard the clipNone; do not redact and keepRedaction cannot make a demonstration of them consented
Agency worker without agency confirmationTreat as declinedHold the clip until confirmation arrivesTheir employer's terms may not allow it
Visitor, customer or driverDiscard or cutRedact if the person is incidental and unidentifiable after reviewNo notice reached them
Audio of anyone not opted inStrip the audio trackDiscard the clipVoices identify people and may raise recording-consent rules

When someone withdraws after filming

Withdrawal has to be as easy as opting in and has to reach every copy. Under the GDPR, a person can withdraw consent at any time, withdrawing must be as easy as giving consent, and withdrawal does not make earlier processing unlawful; that last point is why the cut-off for data already delivered belongs in the agreement rather than in an argument later.

The practical requirement is a pseudonymous operator ID on every clip, derived file and label from the start. With it, the privacy lead can list everything tied to one person in minutes, delete it from site storage and undelivered batches, and send the licensee a removal notice naming file IDs rather than a name. Without it, a withdrawal means reviewing the whole archive by eye.

Keep the key that links operator IDs to names separate, under access control, and out of anything delivered. Licensees should never need it; they need the ID, the scope of that person's yes and the date of any withdrawal.

Worked example: an illustrative stockroom capture

This example is illustrative; the business, counts and timings are invented to show the workflow, not drawn from a real program. A parts distributor agrees to film shelf replenishment and kitting in one stockroom over three weeks.

Fourteen people work the stockroom across two shifts, including two agency workers. After two briefings and a written notice, nine return a yes, three return a no and two answers are pending because the agency has not confirmed its terms. The two agency workers are treated as declined. The scheduler builds capture windows with only the nine opted-in workers in the zone; the five others work receiving during those windows.

In week two, a delivery driver walks through the zone to find a signature. The reviewer finds him in four clips: in three he crosses the background and is cut, and in one he stands at the bench talking to the demonstrator, so the clip is discarded. In week three one operator withdraws by email. Her operator ID returns 11 percent of clips and their hand-pose files; all are deleted from site storage and the next batch before it ships, and the withdrawal log records the file IDs.

The exception: one opted-in worker said yes to video but no to derived hand pose. Her clips stay in the dataset, but the derived file manifest shows no pose files for her ID, and the dataset documentation says so.

What the exclusions take out of the footage

Every exclusion is a sampling decision, so record it rather than hide it. If only the most experienced, most confident workers opt in, the footage will show fewer hesitations, fewer novice errors and fewer left-handed or shorter operators than the real workforce, and a robotics team will train on a narrower slice of the work than it thinks.

Large research corpora show that consented collection at scale is possible: Ego4D reports thousands of hours from hundreds of consenting camera wearers, with de-identification where relevant. A single business cannot match that breadth, but it can report what its exclusions removed. Add a short coverage note to the dataset: how many eligible workers there were, how many opted in, which shifts or tasks have thin coverage and which clip categories were discarded at review. That note is often more useful to a buyer than more footage.

Stop signals that pause the cameras

Some situations mean the program should pause until counsel or the privacy lead has looked at it. Treat these as hard stops, not judgment calls on the floor.

  • A worker reports being asked by a supervisor why they declined, or feels their hours changed because of it.
  • A union, works council or collective agreement covers recording at the site and has not been consulted.
  • The capture zone cannot be separated from customers, patients or the public during capture windows.
  • Minors could enter the zone.
  • The buyer asks for audio, face detail or derived body data that the notice did not mention.
  • Footage would be recorded in a jurisdiction that the scope sheet did not cover, including remote or field sites.
  • The opt-in register and the capture roster disagree, and nobody can say who was in a given window.

Licensing the consented footage later

If the business plans to license this footage, the register, roster, review log and withdrawal log are what make it licensable; they show who agreed to what and that the no answers were honored. For a site that later wants to license demonstration footage, SourceX handles the rights review, contracting, preparation and delivery with the AI developer, and a short eligibility assessment, with no footage uploaded, comes first.

The site's owner approves scope and terms and signs before a single clip is delivered, and blurring or other de-identification is done by SourceX through approved processing routes. Consent records do not replace that review: they answer whether the people agreed, while the agreement separately sets what the business is licensing and to whom.

Frequently asked questions

Can an employer require workers to appear in robot training video?

Making it a condition of the job undermines any claim that the workers agreed, and in the EU regulators say employee consent is rarely freely given. Whether filming can rest on another legal basis depends on the jurisdiction and the facts, so counsel should decide that before any capture. Operationally, the safer design is a program where declining is ordinary and costs nothing.

Is blurring faces enough to include workers who declined?

No. Blurring lowers identifiability but does not turn footage of someone who said no into footage they agreed to, and hands, uniforms and gait can still identify people to colleagues. Cut brief appearances and discard clips where a person who declined is central to the demonstration.

How should agency or contract workers be handled?

Treat them as declined until their agency or employer confirms they may take part and that its own notice reached them. Then give them the same briefing, notice and private decision as employees.

What happens to footage already delivered when a worker withdraws?

The business removes everything it still holds and sends the licensee a removal notice naming the operator ID and file IDs. What the licensee must delete, and how models already trained are treated, is set by the license agreement, which is why it should be agreed before delivery.

Do we need a separate yes for hand pose or body pose files?

It is good practice to ask separately, because derived body data can raise its own rules and some workers who accept video will refuse it. Record it as its own field in the register and generate derived files only for IDs that said yes.

Sources

  • The EDPB says the imbalance of power in employment means employees can give free consent only where refusing has no adverse consequences. Source
  • GDPR Article 7(3) lets a person withdraw consent at any time, requires withdrawal to be as easy as giving consent and says withdrawal does not affect earlier processing. Source
  • The ICO says workers must be told about monitoring before it begins, in an accessible form, and that a DPIA can include discussing plans with workers. Source
  • EDPB video guidelines recommend a warning sign as the first layer of information and say employers should generally not rely on consent for workplace video. Source
  • Illinois BIPA Section 15(b) requires written notice of purpose and length of term and a written release before a biometric identifier is collected. Source
  • New York Civil Rights Law 52-c requires written notice at hiring, acknowledgment and a posted notice for employer monitoring of telephone, email and internet use. Source
  • Ego4D reports egocentric video from hundreds of consenting camera wearers with de-identification where relevant. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies