When an AI-generated asset is questioned, by a client, a regulator or your own legal team, the first request is simple: show how it was made. Many organisations discover at that moment that they cannot.
The usual answer, "we kept the seed", is not the answer it appears to be.
The Seed Is Not the Record
A seed fixes the random starting point of an image. It does not fix the rest of the computation. PyTorch's own reproducibility notes say: "Completely reproducible results are not guaranteed across PyTorch releases, individual commits, or different platforms. Furthermore, results may not be reproducible between CPU and GPU executions, even when using identical seeds."
The Diffusers reproducibility guide reaches the same conclusion and lists the switches needed to reduce variation, adding that deterministic settings can be slower. In short: exact reproduction across machines is possible to approach, expensive to guarantee, and not the default.
What a Defensible Record Contains
| Record | Why governance needs it |
|---|---|
| Model file and version | Different weights, different output and different licence |
| Prompt, negative prompt and settings | Shows intent and method |
| Human inputs: sketches, masks, references | Evidence of human authorship |
| Software version and hardware | Explains why a re-run may differ |
| The original output file | The only perfectly portable artifact |
| Lineage between versions | Shows what changed, when and by whom |
Why This Matters Beyond Engineering
- Licensing: the model version determines which licence applied when the asset was made.
- Authorship: the U.S. Copyright Office ties protection to human input that can be shown, so the inputs are worth keeping.
- Quality control: a lineage lets teams roll back to a known-good version instead of guessing.
- Client trust: "here is the record" ends conversations that "we think it was this seed" starts.
Make the Tool Keep the Record
Manual logging fails under deadline pressure. Records need to be a side effect of the tool. In Vision Studio, our open-source generation app, every render becomes a node in an iteration tree that carries its prompt, model, sampler, steps and seed, records a settings difference against its parent, and survives restarts. We are also explicit about the limit: the engine seeds on the GPU for speed, so a seed is dependable on the machine that made it, not as a cross-machine guarantee.
Who Asks, and What They Need
Different stakeholders ask "how was this made?" for different reasons, and each needs a different slice of the record.
| Who asks | Why | What answers them |
|---|---|---|
| Client | Wants a variation, or proof of originality | Original output, human inputs, settings |
| Legal | Licence compliance, authorship claims | Model and version, licence accepted, human inputs |
| Brand team | Consistency across a campaign | Settings, references, style components used |
| Auditor | Controls exist and operate | Retention policy and evidence it was followed |
A record designed for one audience usually fails the others. A record that captures inputs, settings, model versions and outputs automatically serves all four.
The Cost of Not Keeping Records
Recordkeeping feels like overhead until the first time it is missing. A client asks for "the same image, but in blue" and the team spends a day failing to recreate it. A licence question arrives and nobody knows which model version produced the asset. A dispute arises over authorship and the only evidence is a prompt. Each of these costs more than the storage the record would have taken.
The economics are lopsided. Storing a project's inputs and settings costs almost nothing. Reconstructing them later is often impossible, and the PyTorch guidance quoted above explains why: even a perfect copy of the settings may not regenerate the image on different hardware or software.
Retention Without Hoarding
Keeping everything forever is not a policy either. Most generations are discarded drafts. A sensible rule keeps full records for assets that were published or delivered, keeps lineage for the drafts that led to them, and expires the rest on a schedule. That keeps storage proportionate while preserving the evidence that matters.
Questions to Ask of Any AI Content Tool
- Does it record settings automatically for every generation, or only when someone remembers to export them?
- Does the record include the model file's identity, not just a friendly model name that may point to different weights over time?
- Are human inputs stored with the output: masks, sketches, reference images?
- Can lineage be traced, from a final asset back through the drafts that led to it?
- Can records be exported in a form legal or an auditor can read without the tool?
Start Small
Recordkeeping does not need a platform project. Start with the assets most likely to be questioned: client deliverables, regulated communications and anything carrying a brand's identity. Apply the policy there first, check after a quarter that the records were actually kept, and extend it once it works. A narrow policy that is followed beats a broad one that exists only on paper.
Set Expectations With Clients
Contracts and statements of work increasingly mention AI. Where they do, avoid promising that an asset can be "regenerated exactly" on request. The technical guidance quoted above explains why that promise may not be keepable across hardware and software changes. Promise instead what records make possible: variations built from the retained inputs and settings, delivered files kept for an agreed period, and transparency about which tools and models were used. That is a promise you can keep, which a guarantee of exact regeneration may not be.
A Minimal Policy
- Retain original outputs and human inputs for every published AI asset.
- Record model version and settings automatically, not by hand.
- Note the environment for assets likely to need follow-up work.
- Never promise a client exact regeneration you have not tested.
Governance of AI content starts with a simple test: for any published asset, can you show how it was made? If the answer depends on someone's memory, the record does not exist yet.
The practitioner guide to seeds and reproducibility is on the Vision Studio blog at vision-studio-x.com/blog/your-seed-is-not-a-recipe. Vision Studio is a free, open-source, local-first AI generation app built by Strategia-X. If you want a second pair of eyes on how your organisation adopts AI tooling, talk to us at strategia-x.com.
-Rocky
#AIGovernance #Recordkeeping #Compliance #GenerativeAI #DigitalTransformation #ContentOperations #StrategiaX #VisionStudio
