What’s Your Source File? – Design Week

It is a harder question than it used to be.

The custody problem is old. Every working designer has been called back three years after a campaign to find the source is on a decommissioned server, the fonts were licensed to a colleague who left, and the stock library has been sunset. Losing your ability to reopen your own work has been a design-operations story since the first digital layout, long before any model wrote a prompt. What generative AI has done is make the problem more acute, on more axes, faster.

What has changed structurally is the number of shapes the work now takes. For four decades, “the design” and “the source file” were the same object. INDD, PSD, AI, or a Figma frame: one canonical thing, with derivatives flowing out of it. That equivalence held long enough that the industry built its contracts, DAMs, archives, and handoff practices on top of it.

That equivalence no longer holds cleanly. The source file has become one of several production modes running in parallel, sometimes inside the same job. The infrastructure has not caught up with the range.

Three modes, one designer

Most working designers today move between three workflows, often in a single afternoon.

The first is layered work. The generative operation happens as a layer inside a Photoshop file, an Illustrator artboard, a Figma frame. Generative Fill, Generative Expand, AI masking, AI relighting – these produce pixels that live inside the source, alongside layers the designer has been building all day. The source file remains the canonical object. Ownership, editability, and provenance behave the way they always have, with the caveat that some layers can no longer be reconstructed if the model changes underneath them.

The second is derived work. A designer produces one authored source, and a pipeline derives the campaign’s worth of formatted variants from it. Santa Cruz Software’s BannersUI is one live version of this: the source stays canonical, the outputs are treated as what they are: derivatives. Ownership and editability stay with the source. Provenance carries forward if the pipeline is instrumented for it, which it mostly is not yet.

The third is sourceless work. Prompt to output, no persistent authored file in between. A social tile produced from a sentence at 4pm and shipped at 4.15pm. Ownership, editability, and provenance all have to be reconstructed after the fact if anyone thinks to ask, and often nobody does.

None of these is a future scenario. All three are running in production right now, in the same studios, on the same clients, sometimes by the same designer inside the same brief. The mistake the industry has been making is treating them as a single trajectory. They are three concurrent modes with different infrastructure needs.

The three functions, redistributed

Every design carries three functions that used to be bundled inside the source file: ownership, editability, and provenance. Across the three modes above, they are now distributed differently.

Ownership was tangible when it lived in a file. It still is, in layered and derived work. In sourceless work, ownership splits into two separate questions. The operational claim – can you prove you made this – is substantiable if you retained the prompt, model version, seed, and evidence that you were the operator. Most workflows retain none of these. The commercial claim – can you stop others from using it – is a copyright question, and the answer varies by jurisdiction. In the United Kingdom, the arrangements-maker holds rights under Section 9(3) of the CDPA. Elsewhere, the picture is unresolved and moving. Documentation is one thing. Copyright is another. Studios shipping sourceless work into commercial deployment should know which of the two they actually have.

Editability in layered work is direct manipulation, with a growing footnote about which layers depend on models that may not respond the same way in six months. In derived work, editability lives with the source. In sourceless work, editability is a re-prompt, which is a different operation, filtered through a model whose behaviour is not guaranteed to be stable.

Provenance was carried by the file’s save history, its embedded metadata, and its printer profile if it made it that far. That was a local property of the file. In derived and sourceless work, provenance has to become a portable property – a signed manifest or standardised metadata payload that survives when the asset leaves the environment that produced it. Otherwise the derivative orphans itself the moment it is exported, no matter how well the originating pipeline was structured. C2PA (Content Credentials) is the standards work here, with XMP and IPTC as the older carriers still doing most of the work in practice. Adoption is uneven. The pipelines that do not implement any of it are producing assets that will not be traceable back to their source once they leave the room they were rendered in.

The functions are the same three that always mattered. What has changed is that they no longer come free with a file save.

The infrastructure gap

The rest of the picture is what has not moved.

Contracts still assume a file handoff. Most licensing agreements were written for a world where you can point at the thing you licensed. In derived and sourceless work, the thing is a rendering of a specification, and the specification is not what the contract described. Nobody’s legal team has rewritten this. It is being redefined at the pipeline layer, in silence.

DAMs still catalog files. Whatever gets uploaded, they store. They have no way to know that the “source” being ingested is actually an output, a screengrab of a pipeline run, or a JPEG exported from a sourceless workflow. The catalog looks full. Half of what is in it cannot be reopened.

Archives are the worst of the three, because the failure is temporal. An output archived today from a sourceless workflow may be unreproducible in two years, because the model that produced it has been retired or updated past recognition. Layered and derived work are more forgiving, but only if the pipeline preserved the source and the model dependencies were documented, which is not the default anywhere.

The failures here are the slow kind, the kind you notice five years in, when a client asks to refresh a campaign, and you discover you have the deliverables but not the design.

What to actually do

The honest headline: the source file is now one production mode among several – and most studios, contracts, DAMs, and handoff practices are still built for a world where it was the only one.

The practical response is per-mode.

For layered work, document which layers depend on which models, and re-flatten anything you cannot afford to lose to a model update. Treat model dependency the way you used to treat font licensing.

For derived work, confirm the pipeline preserves the source in a form you can reopen without the pipeline, and confirm the metadata it writes into derivatives conforms to a standard something else can read. If neither is true, you are one vendor decision away from custody loss and one export away from orphaned derivatives.

For sourceless work, decide before the job starts what the asset needs to carry – an audit trail, reproducibility, or neither. Capturing the prompt, model version, seed, and operator gives you an audit trail – evidence of who made what, when, and how – and if the output is signed with a C2PA manifest on the way out, that evidence travels with the asset. Reproducibility is a separate problem, and a harder one. Commercial models drift: weights shift, safety filters update, endpoints get deprecated. True reproducibility requires a locally hosted, version-pinned model, which sits outside standard agency capacity. Ship the asset knowing what you actually retained, and set client expectations accordingly. And know, on the copyright side, what jurisdiction you are shipping into.

None of this is glamorous. It is the design-operations discipline that always did the invisible work of making sure you could reopen your files in five years. The tools changed. The discipline is the same. The studios that update their practices for the three modes will still have their work in a decade. The ones that treat all three as if they were the old workflow will slowly discover that they have deliverables where the design used to be.

Already a member? Sign in.