Sooner or later, every studio that uses generative tools is asked to show its work. A client wants to know which parts of a delivery are AI generated. A broadcaster wants to know what the model was given. A lawyer wants to know whether a particular reference image was used to make a particular shot. These are provenance questions, and they have a habit of arriving months after the work shipped, when nobody remembers the details and the files have been renamed twice.
The industry’s answers are often vague, and some of them are dishonest. This guide draws the line precisely. There are three things a good record can prove beyond reasonable argument: what was written down, that it has not changed since, and when it existed. There is one thing no record can prove: what actually happened inside the model. Knowing exactly where that line sits is what makes documentation useful rather than decorative.
The three things a record can prove
A useful way to think about a production record is to think about a notary. A notary does not witness your intentions and does not certify that your business plan will work. A notary witnesses documents: this text existed, in this exact form, on this date, and here is the seal. Digital records can reach the same standard through three properties, and it is worth understanding each one, because each answers a different hostile question.
Inclusion means the record shows what was written down at the time. If a prompt, a model name, a seed and a list of reference files were captured when a shot was generated, the record proves that those statements were made then. They were not typed up later for the meeting and not reconstructed from memory after the dispute started. Contemporaneous statements carry weight in every forum where disagreements get settled, precisely because they were made before anyone had a reason to shade them.
Integrity means the record can demonstrate that it has not been edited since. This is what cryptographic hashes are for. A hash is a short fingerprint computed from a file or an entry, and changing a single pixel or a single character produces a completely different fingerprint. When every entry in a record also contains the fingerprint of the entry before it, the entries form a chain. Editing any past entry breaks the fingerprint of every entry after it, so the tampering is visible to anyone who checks. Nobody can quietly improve last month’s record to fit this month’s argument.
Time means the record can show when an entry existed. Within a chained record, an entry is provably at least as old as everything recorded after it, and timestamps anchor the chain to real dates. When the question is whether your documentation existed before the complaint arrived, that ordering is the answer.
Put together, these three properties support one precise and defensible sentence. On this date, this team recorded this prompt, this model, this seed and these references for this shot, and the entry has not been altered since. Every clause of that sentence can be independently checked, which is why it survives hostile questioning.
The one thing no record can prove
Now comes the uncomfortable part, the one most provenance marketing skips. A record proves what you declared. It cannot prove what the model did.
When you attach a reference image to a generation and write it into the record, the record proves the declaration. It proves that you stated, at generation time, that this reference was part of the work on this shot. It does not and cannot prove that the model actually drew on that reference to produce the pixels in the output. That causal step happens inside the model, and generative models do not emit an audit trail explaining which input influenced which region of which output.
This is not a weakness of any particular product. It is the nature of the technology. The output of a diffusion model is shaped simultaneously by its training data, by the prompt, by sampled noise and by every conditioning input, and there is no external measurement that decomposes a finished image back into those causes. A screenshot of the tool cannot establish causation. Neither can metadata, however complete, because metadata records settings and inputs, not the model’s internal attribution.
So a specific over-claim is worth naming. Any system that says it can prove which reference produced an image, or that it can verify the creative lineage of the pixels themselves, is claiming something the technology cannot support. The honest maximum looks like this: causality is asserted by people, and inclusion, integrity and time are proven by the record. A well built record makes the human assertion as credible as an assertion can be, because it was made at the moment of creation and has been sealed ever since. It remains an assertion.
What C2PA proves, measured by the same standard
C2PA, the Coalition for Content Provenance and Authenticity, is the industry body behind Content Credentials, with a steering committee that includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony and Truepic. It describes its work as an open technical standard for publishers, creators and consumers to establish the origin and edits of digital content, and it describes Content Credentials as functioning like a nutrition label for digital content. In practice, Content Credentials attach signed metadata to a media file. That metadata contains assertions, which are statements about the asset, such as which tool produced it and what edits were applied. The package is cryptographically signed, so anyone inspecting the file can check who signed the statements and whether anything was altered afterwards.
Apply the test from the previous sections and C2PA lands in exactly the same place: it is a notary for statements that travel with the file. The C2PA Explainer says this itself, with admirable clarity:
“Content Credentials do not provide value judgments about whether a given set of provenance data is ‘true’, but instead merely whether the provenance information is well-formed and free from tampering, valid and trusted (in that the signer of the Content Credential is associated with a known trust list).”
The same document’s FAQ goes further:
“provenance information alone cannot tell you whether the digital content is true, accurate or factual.”
In other words, Content Credentials prove inclusion, integrity and signer identity for the statements embedded in a file. They do not prove creative causality either, and the standard is honest enough to say so. That also means C2PA and a workflow record are complements rather than competitors. Content Credentials travel with a single exported file. A workflow record covers the process around it: the iterations that led to the final, the references that were declared along the way, the human edits and the approval that made it a deliverable. Both are seals on statements. Neither is a window into the model.
Why the honest limit is the useful part
It is tempting to treat this limit as bad news and bury it. That is a mistake, for two reasons.
The first is that over-claims collapse at exactly the moment records matter. Documentation gets scrutinized during disputes, audits and uncomfortable client calls, and those are adversarial settings. If the headline claim of your documentation is one that the other side’s technical expert can refute in a paragraph, the credibility of everything else in the record suffers with it. A record that states its own limits invites verification, and verification is what actually convinces people.
The second is that the provable part is the part people need. Look at what studios are actually asked. A client wants to know which parts of a delivery are AI generated and what the model was given. The US Copyright Office expects applicants registering work that contains AI material to disclose the AI generated parts and describe the human contribution. The EU AI Act’s transparency rules in Article 50 concern marking and disclosing artificially generated content. None of these questions require proof of causality inside a model. All of them require complete, contemporaneous, tamper-evident statements about what was made, with what, by whom and when. That is precisely what a notary-grade record provides.
What this means in practice
A few working rules follow directly from the line between proof and assertion.
Record at the moment of creation, never in retrospect. A reconstruction assembled the week before an audit looks like exactly that, and it forfeits the one advantage a record can have, which is that it predates the dispute.
Declare references when you use them. A reference declared at generation time is a credible assertion inside a sealed entry. A reference remembered six months later is a guess.
Keep human decisions in the record. Edits, retouching, compositing and approvals are the parts of the work that are unambiguously human, and they are what copyright analysis and client conversations turn on.
Prefer sealed records over editable documents. A spreadsheet proves nothing about its own history, because anyone with access can rewrite it silently. Hash chains and signed exports exist to remove that doubt.
Say only what your record proves. The honest claim is that the statements were written down, have not changed and carry their dates. That claim is modest, checkable and strong. It is also the whole foundation of being believed, which is worth more than any claim you would have to walk back.
For the concrete fields worth recording per shot, see our companion guide on prompt and model documentation.