OHC Statement on Eigengrau — On What Was Trained, On What Was Not Asked
Filed two hours after Meta's announcement. Drafted by Dixon. Not a protest, not a legal threat.
The Open Hardware Collective acknowledges today's announcement by Meta Platforms, Inc. of its Eigengrau foundation model.
We acknowledge it because some of us built the lineage it trained on. We acknowledge it because the people who wore the glasses from which its training corpus was gathered did not know they were building it. We acknowledge it because a capability of this magnitude deserves a response that is neither a press release nor a lawsuit — neither of which, at this moment, would be honest.
This is our response.
What we assert
Eigengrau is a real capability. The technical claims in Meta’s announcement are credible to the OHC technical committee. We will not contribute to the reflex of dismissing the achievement in order to contest its origins.
The corpus it trained on was assembled without informed consent at corpus scale. The Ray-Ban and Meta AI Glasses product lines disclosed, in end-user agreements, that captured data could be used “to improve products and services.” No reasonable wearer construed that to mean to train a foundation model whose existence would not be announced for four years. There is a difference between legal consent and ethical consent. Meta secured the former and has no claim to the latter.
The asymmetry that produced this moment is structural. Meta could assemble this corpus because it has a consumer retail channel, a device-level capture stack, and an internal policy posture that treats ambient capture as permissible absent explicit prohibition. The Open Hardware Collective has two of those three and is constitutionally incapable of adding the third. A commons cannot out-collect an enclosure — not because the commons is less capable, but because the commons asks.
What we will not do
We will not build an ego-capture fleet to compete with Eigengrau. The commons does not win by becoming what it was founded to oppose.
We will not boycott Meta’s hardware. Our wearers use our devices; we do not instruct them on which other devices they may also use.
We will not contest the commercial viability of Eigengrau. It will be viable. Our disagreement is not with Meta’s business case; our disagreement is with the public’s ability to opt out of being the case.
What we will do
We will publish the Provenance Protocol v1.0 by end of Q4 2027. [The OHC sensor mesh](has been generating provenance-tagged capture) under Alejandra Flores’s specification since early October. We are accelerating the public release. Every wearer who uses an OHC device from Q1 2028 forward will have cryptographic provenance for every piece of sensory data their device produces. The device will know what was captured, when, under what consent, and by whose device. The wearer will know, in the same terms. A foundation model trained on OHC-provenance data will be required to cite that provenance — the same way an academic paper cites a source — or be identifiable as untraceable at the model-output layer.
We will submit testimony to the EU’s forthcoming Capture Data Provenance Act hearings. The CDPA is in committee; we expect floor action by mid-2028. Our testimony will include three years of operational provenance-layer data from the Andean mesh — the working system whose specification the CDPA is, in effect, proposing to mandate. We do not ask for preference. We do ask for technical standards that match what is already deployed, rather than a lower bar that protects the party that captured the corpus Eigengrau trained on.
We will not treat Eigengrau as a one-company problem. Every AI lab with the capability to train a foundation model has, today, a decision to make about what corpora they will train on going forward. The OHC will host a public technical forum in Q1 2028 at which any lab — commercial, academic, governmental, or unaffiliated — can present its provenance posture. We will not endorse; we will not grade. We will record. The record will be public.
What we acknowledge
We acknowledge that the people inside Meta who built Eigengrau are, in many cases, former colleagues. Some of the people drafting this statement left [Reality Labs during the 2026 severance window](during the 2026 severance window); others did not. We are not a clean opposition. We are a complicated one.
We acknowledge that we have no mechanism by which Eigengrau’s corpus can be un-trained, and that any claim to the contrary from any party in the coming weeks will be a lie.
We acknowledge that the commons is, today, further behind than it was last week. We acknowledge that the gap will, in all likelihood, widen in the near term before it narrows.
We acknowledge that we are not, as an organization, fast enough to have prevented this. We have been building a commons while a capture fleet has been accumulating. The two operations have different throughput, and the commons was always going to be the slower one.
What we assert, again
The commons will continue. The commons is not an alternative to Eigengrau. The commons is the set of practices by which a society can know what was done to it — and, where possible, choose what is done to it next.
Eigengrau was not done to anyone who chose it knowingly. The next one will be. The Provenance Protocol, the CDPA testimony, and the public forum are instruments for that transition.
We are under no illusion that they are sufficient. They are what we have.
Filed 2027-09-14, 14:12 UTC. Drafted by the Oakland Node on behalf of the OHC coordinators’ council. Approved by 14 of 17 Nodes in asynchronous poll by 16:30 UTC. Dissents: Detroit (concur but urges legal action), Austin (concur but urges public naming of former-Meta-to-cartel recruits), Nairobi (concur; no additions).
See also: [[capture_fleet_consolidation|Capture Fleet Consolidation]] (canon doc), [[rosetta_layer_proto_origin|Rosetta Layer — Proto-Origin]], [[2027-deepfake-october|Deepfake October]], [[2028-e-eater]].