Two weeks remain on the EU copyright-and-AI consultation. A robots.txt answer is not a pack for a voice, a face or a camera in someone else’s kitchen.
The targeted consultation is open until 3 November 2026. Today is 19 October. That is a calendar, not a theory. The questionnaire is still the same object it was when the Commission opened it: copyright-protected content used for generative AI, plus neighbouring fights about piracy, research and sound recordings. It asks about an EU opt-out registry, crawler identity disclosure, a rebuttable presumption of use if transparency duties are missed, extended collective licensing, and a possible remuneration right. In parallel, the duties that already apply have not paused for the survey. GPAI providers who place a model on the Union market already owe a copyright policy and a “sufficiently detailed” summary of training content against an AI Office template. The GPAI Code of Practice is how signatories try to show they did it – including crawler measures that start with robots.txt, lawfully accessible sources, and machine-readable reservations of rights.
This is the market. Not a glossary. If you crawl, two weeks is enough time to file and to write down how you honour reservations tomorrow morning. If you buy human data because you cannot scrape faces and kitchens, the same two weeks are useful in a different way. They show, again, that “it was online” was never a rights theory for the modalities that identify a person. Web-copyright is a works problem. A voice, a face, a chest-camera in a home is a person problem. Your counsel should see both. Your dataset card should not pretend they are the same paragraph.
What already applies, whether or not you file
Article 53 is not waiting for the instrument the consultation might produce. Law firms have spent the last year telling providers the training-content summary has to cover pre-training through post-training, live on the company site when the model is placed on the market, and be updated when further training uses new data. The Code’s copyright chapter is the operational version: identify yourself as a crawler, respect robots.txt, treat other machine-readable opt-outs as in-scope once they are state of the art, stay on lawfully accessible material. Measure 1.3 is the sentence about reservations of rights. It is a scrape sentence. It is a good scrape sentence. It does not mention a kitchen.
Aleph Alpha’s Kolibri release earlier this autumn was the product-side signal. A European open-weight model, trained on EU compute, signed up to the Code, sold as something you can run without data leaving the building. Whether that becomes a procurement default is a separate question. The fact they led with training location and Act alignment tells you what enterprise buyers are already asking in RFPs. Two weeks will not invent that question. Two weeks will not retire it either.
Article 10, for high-risk systems, is the quieter cousin and it is still not a copyright form. Training, validation and testing sets are supposed to be relevant, representative, and governed – including the geographical and contextual setting the system will actually run in. That is a data-spec sentence wearing a regulation. A booth speech set for a kitchen product is the wrong object whether or not a regulator is in the room. A US English corpus hosted in Frankfurt is not Irish English in a supermarket headset. Relabelling the card does not change the acoustics.
What a robots.txt filing cannot do
It cannot turn a pile of call-centre audio into consented training data. Copyright opt-outs and crawler rules are about protected works on the web. A voice is a person. Privacy law and the contractual right to train are still different papers. You can comply with a robots.txt regime, publish a model-level summary, and still have no licence to ship a speech model trained on people who rang a shop.
It cannot make research corpora commercial. A Hugging Face card that says “research use” is still the default way those files enter companies. The Act made the question louder. It did not invent it. US acquirers ask it anyway. Two weeks of consultation answers will not rewrite a README that forbade commercial training.
It cannot replace a dataset card. A “sufficiently detailed summary” at model level is not a pack for the human set you added last Tuesday. Training, evaluation, shipping, TTS, faces – those still need yes or no lines next to the object store. A training-content summary on a company site will not tell an auditor whether the kitchen speech was licensed for shipping, whether eval speakers were held out, or whether TTS was even in the ask. Those answers live next to the files, in a pack a person can read. If the pack does not exist, the summary is a press page.
A familiar week in a lab, and it will happen again before 3 November. Someone treats the consultation as the whole rights programme. Counsel drafts the crawler story. Product files a response about registries and remuneration. Meanwhile a speech team is fine-tuning on a “public” podcast dump, a vision team wants Instagram faces with a blur, and a robotics team wants found POV because the hour-counts in the papers look like progress. Blurring is a processing choice. It is not retroactive consent, and it often destroys the signal. Hours of uncleared homes are not a rights-cleared skill set. They are a diligence finding with extra files. The consultation will close. Those files will still be in the weight.
The same week, a prototype that was licensed for research starts to look like a product. Nobody reads the sentence that forbade commercial training. That pattern is older than the Act. The Act just made it harder to shrug. Two weeks is not long enough to invent a new legal theory. It is long enough to stop gluing the web conversation to the person conversation.
What to do with the fortnight
If you crawl: map how you honour reservations of rights, how crawlers identify themselves, and what you would put in a training summary tomorrow. The consultation options – registries, crawler disclosure, presumptions of use – are a preview of where enforcement energy is going. Having the list now is cheaper than reconstructing it from object-store folders in 2027. File if you have a view. Do not wait for the final instrument to start the list. Product teams who wait for the instrument will still need the same records.
If you buy human data: stop treating “ethical” as the pack. Ask who is in the set, what they agreed to, and whether commercial training and evaluation are separately licensed. Internal prototype use is still a purpose. Expanding it later is a new ask. Web-copyright language in a vendor SOW is a warning sign, not a comfort. If the only rights story is robots.txt and “publicly available,” you do not have a voice pack, a face pack, or a home pack. You have a scrape story sitting on biometric files.
If you are in robotics: found first-person video is about to sit in the worst of both worlds – web-copyright heat plus biometric capture. Spec the skill and collect it, or license a set that already has a chain. A wearable in a real home is not an opt-out protocol. The people on the fridge photos did not file a reservation of rights.
If you fine-tune a GPAI model: the summary and the copyright policy are the base model’s problem until you add data. The moment you add a human set, you have a second pack. Do not assume the vendor’s “AI use” line covers shipping, eval, and TTS. Ask for those as separate yes or no answers. The consultation deadline does not extend your licence. It does not shrink it either. It just makes the missing pack louder, because everyone in the building is suddenly talking about copyright.
Keep a written list of what you crawl, what you license, and what you collect. Three columns. Different laws. Different evidence. A voice, a face, a chest-camera in a home will never be answered by a crawler log. Write the use into the brief now – training, evaluation, shipping, TTS, faces – as separate yes or no lines. Expanding a research prototype into a product later is a new ask, not a footnote. If faces are not required, keep them out. If the product is audio-only, do not buy a face-rights problem for a kitchen agent.
The last stretch of a consultation always produces two mistakes. One is silence: wait for the final text, then discover you still need the same records. The other is category error: treat the web-copyright filing as the human-data programme. Two weeks is enough time to avoid both. File the survey if the crawl is yours. Write the pack if the people are yours. Do not let one stand in for the other.
If you are scoping a human set against a use that might include the EU market, write the use into the brief now. The consultation deadline is a calendar. The pack is the work. Training, evaluation, shipping, TTS, faces – as separate yes or no lines – still have to sit next to the object store whether or not Brussels publishes the instrument you hoped for. A robots.txt answer can be finished this fortnight. A voice in a kitchen cannot be finished with robots.txt. That is the distinction the deadline is useful for. Use the two weeks. Do not confuse the forms.
People4.ai will not pretend a scrape is a person. hello@people4.ai · Contact
FAQ
Does filing the consultation finish our GPAI copyright work?
No. The survey is optional feedback on possible future measures. The copyright policy and the training-content summary already apply if you place a general-purpose model on the Union market. File if you have a view. Keep the records either way.
If we honour robots.txt, can we train on voices and faces we found online?
No. Robots.txt and reservations of rights are about works on the web. A voice, a face and a camera in a home are people. You can have a perfect crawler policy and still have no licence to ship.
Is this only an EU problem if our company is elsewhere?
The instrument is European. The buyer question is not. Acquirers and US customers ask where the human inputs came from. If your only answer is “we are not in Europe,” you have not answered them.
Related reading
Brussels is still rewriting who may train on what · What a rights-cleared dataset actually means