Identity without tells.
This is a long-form note on what working identity designers call a tell — the small, structural giveaway that an artefact was produced by a particular kind of tool rather than by a particular kind of hand — and on the specific tell that foundation-model image-generators carry into identity work in 2025 and 2026. The note explains why prompting cannot remove the tell, and what does.
The vocabulary working designers already use.
The word "tell" is borrowed from poker, and it has been in working designers' vocabulary for at least fifty years. A tell is the structural giveaway in an artefact that betrays the tool, the studio, or the era that produced it — usually visible to a practitioner who knows what they are looking for, usually invisible to a layperson. The kerning style of a particular foundry. The grain direction in plate-engraved type. The specific over-correction a phototypesetting machine applied to small caps in the 1970s. Tells are not flaws; they are the part of the work that says where the work came from.
Most working identity designers will tell you, often grudgingly, that their best work has a tell. They want to be readable as themselves to other designers; they want, in a deep sense, to be locatable in the history of the practice. The tell is a signature; it is the part of the work that makes the studio's voice findable.
The problem this note is about is a tell that nobody chose — the tell that arrives in identity work produced by foundation image-generators in 2025–26, and that no client has ever asked for.
A specific, observable thing.
The studio has been calling it the diffusion shadow since mid-2024. It is a specific, observable, repeatable property of identity work generated by foundation diffusion models, and it has at least four components that we have been able to identify.
1. The mean-of-the-internet kerning. Foundation models trained on web-scale image corpora have learnt a kerning rule that is a weighted average of every typeset surface on the visible internet. The result is, statistically, very slightly looser than the kerning a working type designer would set, with characteristic over-corrections at the start of cap pairs — AV, WA, LY. The over-correction is small (typically 1.5–3% of em) but consistent across outputs and across prompts.
2. The compromise terminal. Letterforms in generated marks tend to terminate with a smoothed compromise between the most common terminal shapes in the training distribution — never quite a sharp serif, never quite a clean sans terminal, hovering at a kind of softened wedge that is the visual centre of gravity of the foundation corpus.
3. The over-considered counter. Negative spaces in generated marks are systematically optimised for the readability metrics that the foundation model has internalised — usually by being made slightly more open and slightly more circular than a working type designer would draw them. The result looks legible at every scale and at no scale particularly distinctive.
4. The signature absence of accident. A working mark, even one drawn over six revisions by a careful designer, retains traces of decision — small irregularities, deliberate asymmetries, the mark of the hand that revised it. Generated marks are systematically smooth in a way that reads, to a working designer, as untraceable: there is no decision visible. The smoothness is the tell.
A structural, not a tuning, problem.
The instinct, when one first reads the four components above, is to assume that careful prompting can correct them — specify the kerning, specify the terminals, ask for asymmetry. The studio has tried this, carefully, across roughly 1 200 generations and three months in 2024. The diffusion shadow is robust against prompting in the following structural sense: the prompt steers the model toward subregions of its training distribution, but the foundation model does not contain subregions that lack the shadow. Every subregion of the foundation model's distribution carries the shadow because the shadow is a property of the training distribution as a whole.
This is the difference between tuning and architecture. You cannot prompt your way out of a property that is present in every point of the model's distribution; you can only train into a different distribution that does not have the property. That is what bespoke generative identity work means, and it is the structural reason the studio has refused to take prompt-first commissions.
A different distribution, trained from a different corpus.
The studio's working position is that the only thing that removes the diffusion shadow is training a model on a client-specific corpus that does not exhibit the shadow's four properties. A small generative model — in the 200M to 400M parameter range — trained on a careful corpus of a client's own marks, with proper attention to representing the client's distinctive kerning, terminals, counters, and accident structure, produces outputs that working designers do not read as carrying the shadow.
This is observable and we have tested it. As part of the Identity Loom commission (work file IV.a), we asked seven working identity designers — none of them affiliated with the studio or the client — to look at twenty-five marks and to identify which had been produced by a foundation model and which had been produced by a working designer or by the Loom system. The foundation-model marks were identified at 91% accuracy. The Loom system's marks were identified at 31% — statistically indistinguishable from the working-designer marks, which were identified at 28%. The test is small and we make no broader claim from it; we report it because the studio cares about evidence and because the result has held up across two further informal repetitions.
Why the publisher came to us in the first place.
The publisher in work file IV.a came to the studio in early 2025 with a specific complaint: their junior designers had been using a foundation model to produce drafts of programme covers, and the editorial director could see, on every draft, that the marks had the shadow. The clients of the publisher (older readers, mostly, who had been reading the imprint for thirty years) had not yet complained; the editorial director was confident that they would within a year if the shadow was allowed to settle into the house identity. The commission was a pre-emptive defence of the publisher's voice against a slow drift she could see and could not, before our involvement, name precisely.
The note is in part a written record of what she was seeing — the four components above are her vocabulary as much as ours; she walked the studio through them in a survey afternoon and we wrote them down. The Loom system, when it shipped, was evaluated as much against her ear for the shadow as against any quantitative metric. The publisher's masthead drafts produced through the Loom no longer read to her as shadowed. That is the measure of success.
What this note does not claim.
The diffusion shadow is a 2025–2026 property of widely-deployed foundation image-generators. Future models may carry different tells, or may carry the shadow more weakly, or may carry it in regions of their distribution that are easier to avoid by prompting. The studio expects this and we will revise the note when the evidence changes.
The note also does not claim that foundation models are unsuitable for design work generally. They are excellent at many things — ideation, moodboarding, conceptual exploration, throwaway sketch work where the artefact is meant to be replaced. They are unsuitable specifically for identity work where the output is meant to be readable as the client's for a period of years. The shadow is a marker of provenance; identity work cares about provenance more than most categories.
Where this note draws from.
The four components were named over six months in 2024–2025 by the studio's partners and by three working identity designers we have learnt from. The vocabulary is a reasonable mix of working-designer terms, model-research literature, and the studio's own intake notes. The note does not have a formal academic apparatus — it is a studio position, not a paper — but the reading we have done in adjacent areas (style-transfer metrics, foundation-model bias literature, classical type-design pedagogy) is reflected throughout. If you would like the reading list as a document, write to the studio and we will send it.