Paying a creator to film chores in a home is not bystander consent, and it is not a commercial training pack you can later buy off that person – the camera still sees the environment, other people, children and screens.
Figure’s Index app is the loud version of a pattern every physical-AI lab is now staring at. Pay people by the minute to record real tasks in real kitchens, then train a humanoid policy on the footage. The public numbers are industrial: hundreds of thousands of downloads, millions of videos, a pipeline that ingests half an hour of upload every second. The privacy documents are quieter, and they are the part a buyer or a counsel should actually read. Sensory data is defined as photos, videos or recordings of you and of your environment. Possible disclosures include language that, under some US state laws, can count as a sale. Creators under 18 are told not to register. Homes still contain children. The person who got paid is not the only person in the frame, and they are not the only rights problem in the file.
This is not an argument that Figure should stop collecting. Labs need first-person human video of real work. The papers this year already said so. It is an argument that a creator payout is being asked to do legal work it cannot do. A wage is not a chain. A privacy policy accepted by the filmer is not consent from the partner walking through, the child at the table, the neighbour on the doorstep, or the laptop open on the counter. If you are building a similar funnel, or if you are being offered “Index-like” hours as a commercial pack, that gap is the product risk.
What the documents actually say
Figure’s published privacy policy, updated in January 2026, lists sensory data among the categories it collects: photos, videos or recordings of you and/or of your environment. The “and/or of your environment” is the sentence. The creator is in shot. So is the room. The listed recipients for that category are service providers. Elsewhere the same document says that, depending on state law, some disclosures may constitute a “sale” of personal data, then draws a line at selling or sharing for targeted advertising. Business-transfer language says all personal data it collects may move to a third party in a merger, acquisition, bankruptcy or similar transaction. Retention examples name first name, last name, email, contact-form contents and device records. Video is the heavy object and is not in those examples.
Read that as a product person, not as a slogan. Improving the services includes testing, research, internal analytics and product development. For a robotics company, product development is the policy. Sensory data of a home is how Helix learns to fold laundry. That may be a fair bargain with the adult who applied, received a device, and got paid. It is not automatically a bargain with everyone else the wearable can see. The policy’s children clause is about registration: do not sign up if you are under 18, and the company says it does not knowingly collect personal data from children. A child who is not the account holder does not “send” personal data through a form. They appear in the kitchen while someone else films a chore. A request not to register does not reach that position.
Secondary write-ups have been more aggressive than the primary text, and some of them talk as if commercial purchasers of home video were already a listed recipient row. The document you can actually cite is more careful and still not a pack you can inherit. Service providers. Possible “sale” characterisation under state law. Business transfers. Aggregated or de-identified data that may be disclosed. None of that is “here is a rights-cleared commercial training set, faces out, bystanders out, children out, screens out, licensed for shipping and for eval.” If a vendor later offers you hours that came through a creator app, you are not buying the creator’s consent. You are buying whatever chain that app actually has, which is usually the filmer plus a privacy policy the rest of the household never saw.
Index’s own product page makes the capture setting explicit. Record your day. We come to your home. We come to your business. Cooking, cleaning, laundry, making beds, stocking shelves, busing tables. A second mode books a creator to come and help with chores, which puts a paid stranger and a camera in someone else’s house. That is two consent conversations glued into one upload. The household that booked the help, the creator wearing the rig, the people who live there and were not on the booking, the colleagues in a cafe or a warehouse. Paying minutes does not merge those conversations.
Figure has been in homes as a data strategy for longer than the app brand. Project Go-Big already talked about collecting video of ordinary behaviour at residential scale. Index industrialises the same instinct with a consumer funnel. The rights question did not get easier because the hour-count got larger. Scale multiplies whoever was in the frame.
Paying the filmer does not clear the room
Consented training data is a specific object. People were told what would be captured, what it would be used for, and they agreed in a way you can evidence later. What consented AI training data is is that sentence, not a payout confirmation. A creator who understands they are helping train a robot still may not have been told that the files could sit in a weight, be used for eval, be shown in a demo, or move in a corporate transaction. They almost certainly cannot consent for a partner, a child, a visitor, or a colleague.
Bystanders in homes are not a rare edge. First-person cameras see kitchens because that is where the chores are. They also see faces in the doorway, homework on the table, medicine on the counter, a passport in a pile of post, a laptop with mail open, a television with a news presenter’s face, a neighbour through glass. If faces are not required for the skill, keep them out of the capture. Blur later is a processing choice. It is not retroactive consent, and it often destroys the hands the policy needed. Screens are their own leak. Children are not a “we do not knowingly collect” paragraph when the collection method is a wearable in a family kitchen. You either designed the brief so they are not in frame, with a reject code when they are, or you collected them.
Workplaces are not simpler. A cafe, a warehouse, a factory, an office: other workers, customers, badges, whiteboards, shipment labels. The creator can sign. The people walking behind them did not. “We come to your business” is a deployment sentence. It is also a second population. If you copy the pattern, write who may appear before anyone puts a camera on.
A rights-cleared dataset, for a lab that might ship or get acquired, is boring on purpose. Who is in the set. What they agreed to. Which uses are actually licensed: training, evaluation, shipping, marketing stills, TTS if there is audio, faces as a separate yes or no. What a rights-cleared dataset actually means is that pack sitting next to the object store. A creator-app privacy policy is a notice to users of an app. It is not that pack. It will not tell an auditor whether bystanders were in scope, whether children were a reject, whether commercial training by a third-party lab was in the ask, or whether eval copies may be kept and not trained on.
Here is the in-house version, without Figure’s brand. A robotics team wants hours. A vendor says they can stand up a creator network this quarter. People film chores on a phone or a supplied rig. QA is “the video uploaded.” Counsel asks who is in the files. The answer is “the creators consented.” A diligence reviewer opens five clips. A child at a table. A partner’s face. A work badge. A WhatsApp thread on a laptop. Friday’s hour-count looked like a data factory. Monday the pack does not exist, and you cannot unfilm a home.
Europe does not save you with a crawler story. The copyright consultation and the GPAI training-content summary are about works and sources at model level. A chest camera in a kitchen was never “it was online.” Article 10, if you are in a high-risk box, asks whether the sets match the setting. Homes are a setting. Children in those homes are not a diversity slide. Counsel in the Union will not treat a US app privacy policy as consented training data for everyone in the room. US acquirers will ask the same question with different statutes. The brand on the app does not answer it.
What a lab should write before it copies the funnel
Start from the skill, not from the app. “Fold a towel on a bed, chest camera, indoor daylight, faces out, no screens, no other people, start to finish including the failures.” That sentence decides the reject codes. Task not completed. Camera on the ceiling. Non-consented faces. Children. Screens. Beauty filters. Digital zoom that destroys the hands. Automation can flag black frames. A person still has to apply the codes that mean “this is not the job.” If gold items do not exist before production, the first month is the network learning your kitchen on your budget.
Split eval by person so the same hands are not in train and test. A creator who filmed four hundred minutes is one person, however many tasks they touched. Random clip splits are how a policy memorises a household. Hold those people out. Rights for training and for keeping the eval set are separate lines. A creator who agreed to “help train robots” has not automatically agreed that their kitchen is a frozen exam you re-run for years, or that a still from the take can go on a launch site.
If faces are not required, keep them out at capture. If other people can walk through, you do not have a home-skill set. You have a biometric-scale collection with extra chores. If you need workplaces, get a workplace agreement that covers the people who work there, not only the person wearing the rig. If a catalogue item already matches the skill and the pack exists, look at it with the actual trainer before you stand up a consumer funnel. Custom when the object set, the kitchen, or the faces decision is yours.
Do not buy the hours off the creator later and call that a commercial pack. The creator can assign what they have. They cannot assign the bystander’s face, the child’s voice in the next room, the screen, or a purpose they were never told about. If your use is shipping a policy, eval, or a third-party licence, that use had to be in the ask. Expanding a “research and product development” privacy notice after the weight exists is how teams discover they have a diligence finding with extra files.
The market will keep rewarding hour-counts. Index exists because found internet video is the wrong viewpoint and the wrong rights theory, and because buying from vendors did not hit the throughput Figure wanted. That is a real constraint. The correction is not “stop collecting in homes.” It is “a home is a rights environment.” Pay the adult. Write who else may appear. Reject the takes that break the sentence. Keep a pack you can show. Paying a creator is how you get the skill into the camera. It is not how you clear the house.
If you are scoping first-person capture in homes or workplaces this quarter, the brief is the dataset. Who is in frame. Who agreed. What happens when a child walks through. Whether screens are a reject. Whether a later commercial licence is even on the table. An app can be the funnel. It cannot be the pack.
People4.ai will not pretend a payout is a household’s consent. hello@people4.ai · Contact
FAQ
If the creator got paid and tapped accept, is the home cleared?
No. Their agreement covers them, and only for the uses they were actually told about. Bystanders, children, visitors and screens are different people and different objects. A wage is not a chain for the rest of the room.
Can we blur faces and screens after upload and call it done?
Blur is a processing choice. It is not retroactive consent, and it often destroys the hands or the object the policy needed. If faces and screens are not required, keep them out of the capture and reject takes that include them.
Could we license Index-like footage from creators as a commercial training pack later?
Only for rights they actually hold. They cannot sell you the bystander, the child, or a purpose that was never in the ask. If shipping, eval or sublicence were not in the original agreement, you are negotiating a new object, often an impossible one.
Related reading
What is consented AI training data? · What a rights-cleared dataset actually means