AI-mediated communication systems measurably shift the opinions of the groups they serve, and the question “did this person say this thing” is becoming harder to answer. Durable content provenance requires three layers working together: C2PA for structured manifests, blockchain for platform-independent anchoring, and decentralised identity to bind the signer.
The published research is already there. AI-mediated communication systems, the kind that sit between a writer and a reader and offer polish, suggestion, summarisation, or rewrite, measurably shift the opinions of the resulting groups. Not in catastrophic ways. Not necessarily in directions anyone chose deliberately. But systematically, observably, and reproducibly. When LLMs sit between authors and audiences, the median opinion of the audience moves toward something the LLM is doing.
This is not an accusation against any specific platform. It is a description of a category. Every social platform that helps users write posts is becoming an AI-mediated communication system. Every messaging client that offers smart replies is one. Every email tool that suggests “write this in a more professional tone.” Every comment box with a polish-the-grammar feature. Every assistive context note. The category is expanding faster than the discussion of what its effects are.
The result, taken as a whole, is that the question “did this person say this thing” is becoming meaningfully harder to answer than it used to be. Not because identities are forged. Because the path from “what the human meant” to “what arrived on the platform” now routinely runs through a model that nudges, rewrites, or replaces words along the way.
The architectural response to this problem has been quietly maturing for a decade. It is called content provenance, and the dominant standard for it is the Coalition for Content Provenance and Authenticity, or C2PA, with a complementary adoption community in the Content Authenticity Initiative. This piece builds on Friday’s argument that reputation is becoming public infrastructure, extending the same insight from people to content.
What C2PA does well
C2PA attaches a cryptographically signed manifest to a piece of content. The C2PA Technical Specification 2.2 records who created the content, when, with what tools, whether AI was involved at any step, and what edits have been made since capture. Any change to the manifest, or any tampering with the content, breaks the signature. The signature is verifiable by anyone with the public key of the signer.
In its design, C2PA is the right answer. Cameras can sign captures. Editing tools can append edit records to the manifest. Publishing platforms can validate the chain on ingest. Readers, in principle, can inspect the manifest and see what the content has been through.
The problem is the layer C2PA lives at.
Where C2PA gets fragile
C2PA manifests are bound to specific files. The file moves through the network. The manifest is metadata attached to the file. When platforms strip metadata for legitimate reasons (privacy, file size, compatibility) the manifest goes with it. When platforms re-encode or re-render content (most do, routinely) the manifest must be re-applied by the platform, which means trusting the platform to do so faithfully. When the file is screenshotted, quoted, recombined, or rebroadcast, the manifest is lost.
The standard is correct. The application layer is not always honest, and the file format is not always durable. Provenance that survives a single hop along a friendly path is good. Provenance that survives an arbitrary number of hops along arbitrary paths is what the AI-mediated discourse environment actually needs.
The missing piece is durability that does not depend on whether the next platform in the chain bothers to preserve it.
What blockchain adds
A blockchain is, among other things, a public, append-only, timestamped record. Anyone can write to it. Nobody can rewrite or delete what is already there. The records are verifiable by anyone, do not depend on any single platform’s good faith, and survive arbitrarily many hops because they are not attached to the file at all. They are attached to the content’s cryptographic hash, which any honest verifier can recompute from the file itself.
Pair a C2PA manifest with an on-chain anchor and the manifest becomes durable. The C2PA layer continues to do what it does well: structured edit history, tool attribution, AI disclosure, signing semantics. The on-chain layer holds the cryptographic anchor that lets anyone, at any future point, verify that the manifest existed at a specific time and was bound to specific content. If the file’s manifest is stripped, the anchor lets you reconstruct the chain of custody from the content itself. If the file is re-encoded, the hash anchor lets you confirm whether the new file matches the originally anchored content.
The two layers do different work. C2PA describes the journey. The chain attests, for as long as the chain exists, that the journey was described in a particular way at a particular time.
Identity is the part that ties it together
The chain anchors a hash. The hash binds to a manifest. The manifest names a signer. The signer needs a durable, portable, verifiable identity, or the whole structure points at nothing.
This is the part of the architecture that decentralised identity standards complete. The signer of a C2PA manifest is identified by a DID. The DID resolves to a public key. The public key verifies the signature. The DID can be controlled by an individual, an organisation, a publication, a device. The DID outlives any specific platform. The DID’s credentials that attest to the signer’s identity describe who the signer is in terms the verifier can trust without contacting the signer. (This is the same primitive Tuesday’s foundational piece introduced for evaluators; it is the same primitive here.)
In this architecture, “did this person say this thing” becomes a question with a tractable answer. The reader, or any tool acting on behalf of the reader, can verify the content’s hash against the chain anchor, confirm the manifest’s chain of custody, validate the signer’s DID, and check the credentials that attest to the signer’s identity. None of these steps depend on any single platform staying honest. None depend on any specific file format surviving.
Why this matters now
The volume of AI-mediated communication is growing faster than the deployment of provenance infrastructure that survives platforms. The window in which “I said this and here is durable proof” remains a feasible architectural claim is narrowing. C2PA on its own is necessary but fragile. Blockchain anchoring on its own does not capture the structured provenance C2PA provides. Decentralised identity on its own does not bind to specific content. The three together are the architecture that does.
Ontology has been deploying decentralised identity standards for years. Pairing them with content provenance standards and chain-based anchoring is the natural next move for any team that wants their content, their attribution, and their authorship to remain durable in an environment where every platform between author and reader is becoming an AI mediator.
The default outcome, if no one builds the integrated architecture, is that “who said what” becomes a soft, platform-dependent claim. The alternative is to build, now, the infrastructure that keeps it hard.
Continue reading this week
Tomorrow: DIDs for agents: identity for the next class of actors, the forward-looking close on what comes next.
