The EU Code of Practice can require crawlers to honour robots.txt, lawfully accessible sources and reservations of rights, and you can do all of that and still have no licence for a chest camera in a home.
That sentence is the gap counsel keeps flattening into one “training-data policy.” The General-Purpose AI Code of Practice, copyright chapter, is a crawler document. Signatories who scrape the World Wide Web for text and data mining commit to employ crawlers that read robots.txt as specified in RFC 9309, to identify other machine-readable reservations under Article 4(3) of the DSM Directive, to reproduce only lawfully accessible content, not to circumvent paywalls, and to stay off sites recognised as persistent commercial-scale infringement. Article 53(1)(c) of the AI Act already required a copyright policy that identifies and complies with those reservations. The Code is how a lot of labs evidence it. Fine. A chest camera in a kitchen is not a URL. There is no robots.txt on a hallway. A reservation of rights in a site header does not reach a child at the table.
The market this autumn is full of people pointing at crawler hygiene as if it were a capture licence. GPAI training-content summaries, opt-out protocols, a complaint mailbox for rightsholders. Those are real obligations if you place a general-purpose model on the Union market. They are about works on the web. Faces, voices, first-person rooms and biometric capture in homes were never that category. You can run a perfect crawler and still have no chain for the wearable.
What the Code actually covers
Read the measures as a product person, not as a slogan. Draw up a copyright policy for GPAI models on the EU market. Extract only lawfully accessible copyright-protected content when crawling. Honour machine-readable opt-outs, starting with robots.txt, then other protocols that make it through a standards or EU-facilitated process. Mitigate infringing outputs. Name a point of contact. That is the chapter. Compliance with the chapter helps evidence the Article 53 duty. It does not, on the Commission’s own framing, guarantee conformity with copyright law, and it does not mention a kitchen.
“Lawfully accessible” in this chapter is a web sentence. Do not break a paywall. Do not treat a logged-in corpus as a public crawl. Do not hoover sites a court has already called a piracy shop. It is not “the householder opened the door.” A creator who let a camera in has made the room physically accessible. They have not made the faces in it lawfully available for TDM, and TDM is the wrong statute for most of what a chest camera picks up anyway. Copyright opt-outs protect works. They do not enrol a bystander.
robots.txt is a file on a host. RFC 9309 tells a well-behaved crawler where not to go. Households do not publish one. Neither do the partner in the doorway, the child, the neighbour, the laptop, or the television presenter whose face is on in the background. Honouring robots.txt is what you do instead of pretending the open web is a licence. It is not what you do instead of a capture brief. If your physical-AI pipeline’s compliance memo starts and ends with crawler settings, the memo is about the wrong collection.
The other protocols the Code nods at – asset-level metadata, location-based flags, whatever a later standards process blesses – are still flags on content made available online. A wearable recording of “fold a towel on this bed” is not content the rightsholder posted with a machine-readable reservation. It is a recording you caused to exist. The legal objects are consent, purpose limitation, biometrics, children, employment if it is a workplace, and whatever commercial licence you actually asked for. Mixing those into a copyright policy because the Code is on the intranet is how teams fail an audit with a well-documented crawler.
Brussels is still rewriting who may train on what. That consultation is the web and the work. Robot teams using found human video have a version of the same problem with a biometric twist. Counsel will not treat a kitchen as “it was public,” and they will not treat a signed robots.txt log as a substitute for consented training data.
A home is not a URL
Physical-AI capture in 2026 is first-person on purpose. Chest, head, glasses. Chores, shops, warehouses. Figure-style creator funnels, lab-run pilots, vendor collections. The camera sees the skill because that is the viewpoint the policy will have. It also sees the environment, other people, children, screens, post, medicine, badges. Paying the filmer, or honouring every robots.txt on the internet the lab also crawled that quarter, does not touch those objects.
Consented training data is a specific pack. People were told what would be captured, what it would be used for, and they agreed in a way you can evidence later. The enrolled adult can agree for themselves, for the uses in the ask. They cannot robots.txt-opt-in the rest of the house. Bystanders in homes are not a rare edge. First-person cameras see kitchens because that is where the chores are. Workplaces are a second population: colleagues, customers, whiteboards, shipment labels. “We come to your business” is a deployment sentence. It is also not a crawl.
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, faces as a separate yes or no. That pack sits next to the object store. A copyright policy that says the crawler obeys RFC 9309 will not tell an auditor whether children were a reject, whether screens were in scope, whether commercial training by a later acquirer was in the ask, or whether eval copies may be kept and not trained on. If you cannot point at those lines, you do not have a cleared home set. You have a crawler story and a pile of kitchens.
Here is the in-house version, without a brand. A robotics team wants hours. Counsel asks about the AI Act. Someone pulls the GPAI Code and the robots.txt logs. The capture is a chest camera in rented houses. Diligence opens five clips. A child at a table. A partner’s face. A work badge. A WhatsApp thread on a laptop. Friday’s compliance memo looked like Measure 1.3. Monday the pack does not exist, and you cannot unfilm a home. The Code did what it said. It never claimed to cover the wearable.
Article 10, if you are in a high-risk box, is closer to the right instinct and still not a crawler. It asks whether the training, validation and testing sets are relevant, representative, and as free as possible of the errors that would break the use. Homes are a setting. Children in those homes are not a diversity slide. A set that matches the setting is a capture spec: who may appear, what the camera may see, what a reject is. Honouring reservations on the web does not make a home set representative of a home. It makes a web set less rude.
US acquirers will ask the same question with different statutes. Biometric laws, children’s codes, employment rules, “sale” language in a privacy policy. The brand on the crawler does not answer it. Neither does “the content was lawfully accessible” if the content is a person who was never asked.
Write the setting, not the crawler policy
Start from the skill and the room. “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 person who agreed to “help train robots” has not automatically agreed that their kitchen is a frozen exam you re-run for years.
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.
Keep the crawler policy. You may need it. Label it as the web chapter. Put a second chapter next to it for anything a person wore, filmed, or said in a room. Who is in frame. Who agreed. Purpose. Children. Screens. Shipping, eval, stills, sublicence. Retention. What happens in a business transfer. If you cannot write that chapter without quoting robots.txt, you are not ready to put a camera on a chest. You are hoping a copyright code will stretch over a biometric capture. It will not.
The market will keep rewarding hour-counts and crawler hygiene, because both look like seriousness in a slide. Hour-counts without a household’s consent are a rights problem at industrial scale. Crawler hygiene without a capture brief is a different dataset’s homework. Found internet video is the wrong viewpoint and the wrong rights theory for a home skill, which is why labs went into houses in the first place. Going into houses and then quoting the Code of Practice is how you bring the wrong theory with you.
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. Honour robots.txt on the web you still crawl. Do not confuse that file with a home. A reservation of rights is what a publisher puts on a work. A chest camera is what you put on a person. Those are not the same licence.
People4.ai will not treat a crawler log as a household’s consent. hello@people4.ai · Contact
FAQ
If our GPAI crawler obeys robots.txt and paywalls, are our robot videos in the clear?
No. Those commitments cover web collection of copyright works. A chest camera in a home is a different capture, with different people and different statutes. You still need a chain for everyone in frame and every use you will actually make.
Can we rely on the householder’s accept and the Code of Practice together?
The householder can agree for themselves, for the uses they were told about. The Code does not enrol bystanders, children, screens or a purpose that was never in the ask. Two incomplete documents do not make a pack.
Does Article 10 fix this if we are high-risk?
Article 10 asks whether the sets match the setting. It does not replace consent, and it does not make a crawler policy into a home licence. Spec who may appear, reject the takes that break that sentence, and keep a pack you can show.
Related reading
Brussels is still rewriting who may train on what · What is consented AI training data?