Provenance is now a spec: the EU AI Act and robot data

A robotics team wants to buy 500 hours of manipulation data. The purchase contract lands on an engineer's desk with one clause that was not there a year ago. For every clip, the seller must supply consent records, sensor calibration, and an unbroken chain of custody from capture to delivery. The seller has the footage. The seller does not have the log. The price falls.

This is what the EU AI Act does to training data, and most teams still read it as a legal problem when it is an engineering one. Regulation (EU) 2024/1689 does not ask you to file paperwork after the fact. It asks you to prove where your data came from. Proof is something you either recorded while the sensor was running, or you did not.

That distinction is the whole post. Provenance does not survive being reconstructed. You can estimate a hand pose from pixels. You cannot estimate whether the person in frame consented, what the depth camera's calibration was that afternoon, or which annotator labeled which clip under which guideline. Those facts exist for one moment, at capture, then they are gone unless something wrote them down.

The provenance you did not record is gone

Text models got lucky. The internet is its own audit trail: a page has a URL, a timestamp, an author, and a license you can at least argue about. Egocentric robot data has none of that exhaust. A person puts on a capture rig, does a task once, and the moment passes. If the rig did not stamp who consented, when each frame was taken, and how the sensors were set, that context is unrecoverable. Nobody can re-derive it from the video alone.

Consent is the sharpest example. A recording of someone's hands, home, and face is personal data, and the European Data Protection Board treats it as such no matter what you later train on it. If you did not capture a consent record tied to that person and that session, you cannot manufacture one afterward. Once faces are blurred and the raw is deleted, you have also deleted your ability to prove the recording was lawful. The clean-looking anonymized dataset is the one you can no longer defend.

The same logic runs through every provenance field. Calibration drifts during a session and is either measured or guessed later. Channels are stamped to a shared clock at capture, or they are re-aligned by hand with offsets you can never fully recover. Each annotation carries the annotator, the guideline version, and the time, or it is a flat label file with no memory of how it was made. In every case the honest record costs almost nothing at capture and cannot be bought at any price after.

Article 10 turns data quality into a documented process

Read the Act's own framing of high-risk systems and Article 10 stops sounding like compliance and starts reading like a data-pipeline spec. It requires that the training, validation, and testing datasets behind such a system be produced under data-governance practices that cover the collection process, the origin of the data, and the annotation and labeling procedures. In plain terms, it wants a documented account of how the data was made, not just what it contains.

For a robot dataset that maps onto concrete build decisions. The origin of the data means a chain of custody. The collection process means logged capture conditions. Annotation procedures mean versioned labeling guidelines and a record of who applied them. None of this is exotic to a team that already treats data as an engineered artifact. The Act simply makes the artifact non-optional and its documentation inspectable.

Provenance you can reconstruct was never really provenance. It is a story you tell after the fact. The Act asks for the record you kept while the sensor was still running.

The European Parliament's plain-language explainer is blunt about the direction of travel: high-risk systems carry documentation and traceability duties, and enforcement phases in across 2026 and 2027 rather than arriving all at once. The exact application dates have moved during implementation, so treat the timeline as a slope, not a single cliff. The practical reading does not change. Data you capture in 2026 is data you may have to document in 2027.

Captured at source versus bolted on later

The gap between the two ways of getting provenance is not a matter of effort. It is a matter of what is still knowable. The table below takes six fields a buyer will ask about and shows what each becomes when you try to add it after capture instead of at the sensor.

Read the right-hand column as a list of things a diligence process will find. Every gap is a discount, and some gaps are disqualifying. A dataset that cannot prove consent is not cheap, it is unusable for a system headed to the EU market.

Provenance becomes a line item in procurement

Here is the commercial turn that engineers underrate. When a robot OEM buys training data for a system it intends to sell in Europe, it inherits your provenance. Its own Article 10 file is only as strong as the manifest you hand over. So the buyer's counsel starts asking for that manifest before signing, the same way a chip buyer asks for a datasheet.

The manifest is a concrete deliverable, not a vibe. For each clip it answers: who consented and under what terms, how the raw was anonymized and whether it can be deleted on request, how the sensors were calibrated, how the channels were synchronized, who annotated what against which guideline, and what license governs downstream use. A dataset that ships with that document commands a premium. A dataset without one is a liability the buyer has to price in, or walk away from.

Industry coverage in outlets like The Robot Report already tracks humanoids moving from lab demos toward commercial pilots. As those systems approach the EU market, the diligence on their training data tightens in step. The dataset that wins is not the largest. It is the one whose provenance a lawyer can sign off on in an afternoon.

A capture-time spec, like frame rate

No serious team decides frame rate after recording. You set it before you press record, because you cannot add frames that were never captured. Provenance now belongs in that same category. Consent capture, calibration logging, time-sync, and annotation lineage are parameters of the capture session, fixed at design time, not features you sprinkle on during cleanup.

The GDPR reinforces this from the other side. Because egocentric footage is dense with personal data, lawful basis and retention limits apply from the first frame, and the AI Act explicitly leaves those obligations intact. A pipeline that logs consent and provenance as structured metadata is already most of the way to satisfying both regimes at once. A pipeline that defers them is accumulating debt that compounds the day an auditor or a buyer arrives.

So the useful way to hear the EU AI Act is not as a rulebook to survive but as a specification to build against. It says, in effect, that a dataset is only as valuable as the record of how it was made. That record has to be written at the one moment it can be, while the sensor is still running. Provenance is not paperwork you attach later. It is a capture-time spec, and the teams that treat it that way will be the ones with data left to sell.