Imagination is still a superpower

The professional grief cycle around AI has a specific texture. David Kessler, who added a sixth stage to Elisabeth Kübler-Ross's original five in Finding Meaning (2019), described bargaining as the phase where resistance starts looking like adaptation. The designer who once said "AI can't replace real creative judgment" now says "I'm repositioning as an AI-augmented designer." The consultant who protected their proprietary frameworks now publishes them as prompts. The strategist who built their reputation on synthesis now calls themselves a "sense-making layer." The motion is real — genuine tools are being adopted, genuine workflows are being rebuilt — but the underlying question, what remains once the machine absorbs the convergent layer of creative labour, tends to go unasked. Asking it directly feels like standing too close to an edge that is still approaching.

The failure mode of bargaining is specific: it generates motion that looks like resolution without being it. A methodology is not a position. A LinkedIn post about AI-native workflow is not the same as doing it. The designers most comfortable right now — those who have integrated AI most visibly, who write and speak about the transition — may also be the ones who have been most successful at avoiding the question at the centre of it.

What follows is an attempt to ask it directly. The argument has five parts. First, what designers should hand to the machine. Second, the specific physics of AI-assisted design work, and why a recent shift in tooling doesn't change as much as it appears to. Third, why the distinction between taste, vanguard, and imagination matters practically rather than philosophically. Fourth, why imagination, understood precisely, has become more load-bearing than it has ever been. And fifth, the reality problem the whole argument is downstream of: that synthetic imagination drifts from culture's first-order material, which makes the human attached to that material newly central.

I. Stop defending the hygiene

A significant portion of what designers have historically been paid to do is coverage: confirming the accessible navigation tree is sound, the error states handled, the responsive breakpoints consistent, the checkout flow has a return path, the form validation meets expected patterns, the loading and empty states accounted for. This work is not trivial — it requires real knowledge and carries real consequences when it fails — but it is, structurally, a question of whether a given output meets established standards. That is a pattern-matching problem, and large language models are exceptionally good at pattern-matching against documented standards.

When a designer prompts an AI agent to build an e-commerce site for a mug brand, the machine produces something with inventory logic, a return flow, accessibility attributes, and responsive breakpoints. It produces things the designer had not yet articulated as requirements, because those requirements are well-represented in the training data, drawn from thousands of similar implementations. The AI is not being creative; it is retrieving a statistical consensus of what a functioning e-commerce site requires. But the retrieval is genuinely useful, in the same way a thorough checklist is useful: it closes the gap between what you thought to specify and what the problem actually demands. The designer who once spent the first third of a project building and verifying this infrastructure now receives it as a starting condition. That is a transfer, not an augmentation. Work that existed is gone; capacity that was consumed is available.

The denial version of this story sounds reasonable from the inside: AI can assist with the basics, but the real design judgment is still mine. You add an AI layer to your existing process, produce the same deliverables faster and cheaper, and describe it as transformation without seriously examining what has actually transferred. The tell, looking back, is that the core of what you do never gets examined. The question worth asking — what was the coverage work costing in terms of the attention it consumed — tends not to get asked, because its answer leads somewhere unsettling.

Most designers, honest with themselves, will recognise that a significant portion of their cognitive load on any given project was allocated to precisely the work the machine now handles adequately. The research sweep that confirmed what you already suspected. The accessibility audit that flagged the seven issues you knew were there. The pattern library search that found the component you were going to find anyway. None of this was wasted time — it was the work, and doing it built the competence that made you trustworthy. But the attention it consumed was not, in retrospect, being spent on the thing that makes a design project worth doing. That attention is now available. The question is whether it gets redirected toward the work the machine cannot yet do, or quietly absorbed into a faster version of the same process.

What that redirection looks like depends on a distinction — between work that has a correct answer and work that does not — that the rest of this essay tries to draw precisely enough to be useful.

II. The command line, the canvas, and what Figma just shipped

The interface through which a designer interacts with an AI agent is not neutral. It shapes which decisions can be made in which order, and therefore which capabilities remain valuable.

The classical AI workflow — Claude Code, Cursor, the terminal-adjacent agents that defined the first phase of AI-assisted design — operates on a command-line model. The designer writes a prompt. The machine processes. The output reloads. Tokens burn. And crucially, the designer must commit to a direction before seeing its result. The visual judgment that in a graphical tool happens mid-gesture — dragging an icon across a composition, reading the imbalance before the hand makes its final move, adjusting without ever fully stopping — here has to happen entirely inside the designer's head before the first word of the prompt is written.

The shift is not minor. It is a change in where cognition is allowed to live.

What graphical tools allowed

• navigate before fully knowing
• discover through manipulation
• distribute cognition across hand, eye, and screen

What command-based AI requires

• specify before seeing
• hold the image internally first
• commit direction before feedback arrives

Figma's design agent, launched in May 2026, is best understood as an explicit attempt to address this mismatch. The announcement framed the problem directly: speed or precision, AI generation or direct manipulation — you shouldn't have to choose. The agent works on the multiplayer canvas, aware of components, variables, and the standards a team has built over time. Designers and agents work in parallel on the same surface. A designer at Notion described doing "almost all of my design ideation on the canvas" and then having the agent "take it the last mile."

The advance is real. It should not be granted more conceptual dignity than it deserves. What Figma's agent reduces is friction inside execution. It does not solve the epistemological problem of direction. The designer still commits before the output exists. The dependency on a destination in mind before the journey begins is not removed; if anything, it is deepened, because iteration becomes so cheap that motion can be mistaken for vision.

Faster iteration risks producing a new form of professional confusion: the belief that acceleration itself constitutes direction. Cheap iteration can conceal the absence of a coherent image.

The agent is well-suited to convergent problems — apply the right component, maintain the grid, generate responsive variants, run the accessibility pass, produce ten colour-scheme iterations. These have correct answers within existing standards. A designer who spends time navigating convergent problems manually in 2026 is, in a real sense, spending time they do not have to spend.

The work that does not have a correct answer yet — what the mug brand should feel like, which visual grammar is worth drawing from, what in the current culture it should be in conversation with, and how that conversation should be made visible — is a divergent problem. The command-line and canvas-agent paradigms are useful here only to the extent that the designer already knows what they want to produce. How to know what you want is a different problem, and it is the one the rest of this essay addresses.

III. Taste is the wrong word

Ask any of the major AI systems for guidance on design quality and you receive a version of the same answer: develop your taste, exercise your judgment, cultivate a point of view. The advice is not wrong. But the AI systems are not themselves performing what they recommend, and the gap between the advice and the performance reveals a distinction that matters practically.

"Taste," "vanguard," and "imagination" tend to be used as if they were points on a single scale. They are three different operations with very different relationships to what AI can and cannot do.

Taste, vanguard, imagination

• Taste identifies coherence inside an existing grammar
• Vanguard breaks the grammar before culture absorbs it
• Imagination holds the next image before it exists externally

Taste, in the technical sense Umberto Eco develops across Opera Aperta (1962) and Apocalittici e Integrati (1964), is not a personal quirk or intuitive preference. It is a trained capacity to identify coherence, discrepancy, and exaggeration within an established aesthetic grammar. Eco's taxonomy of kitsch, midcult, and high culture turns entirely on this distinction. Kitsch is not a style; it is an operation — the prefabrication of effect, the deployment of borrowed formal moves to produce a predictable emotional response without the interpretive effort the original demanded. Midcult is more insidious because it is more ambitious: it preserves the surface markers of seriousness — the critical vocabulary, the formal complexity, the cultural references — while eliminating the actual difficulty. It flatters the audience into believing they are doing the work while doing it on their behalf. High culture, for Eco, is distinguished not by complexity for its own sake but by the refusal to complete the interpretation: the work opens a problem rather than closing it.

What AI systems can do, with increasing sophistication, is operate within established grammars. A well-prompted model can maintain the internal coherence of Swiss Modernism, reproduce the grid structures of Dutch editorial design, apply the visual conventions of contemporary fintech, or produce packaging that meets the expectations of the artisanal food market. In Eco's terms, this is kitsch production in the technical sense: prefabrication of effect through borrowed convention. Not an insult — kitsch production is legitimate and economically significant, and most commercial design is kitsch in Eco's sense without that being a problem. The mug brand needs a visual identity that communicates artisanal quality and contemporary sensibility; the model produces something that does this efficiently and competently, drawing on the formal vocabulary the market has already established as the carrier of those values.

Selecting which grammar to apply is itself a probabilistic calculation: given the brief, the audience, the competitive context, which conventions carry the highest likelihood of being received as appropriate? Recommendation systems have been performing the adjacent version of this for years, not by generating taste but by reinforcing it. They surface content and creators whose vocabulary is likely to influence the people who encounter them, tightening the feedback loops through which individual preferences become shared conventions, which then become market expectations, which then become the constraints within which the next generation operates. Algorithmic reinforcement does not produce new aesthetic positions. It accelerates the consolidation of existing ones.

Vanguard is a different operation, and the difference is not one of degree but of kind. The avant-garde gesture, as Eco understood it and as Doshi and Hauser (2024) document empirically in the context of AI-driven creative homogenisation, does not draw from the established repertoire. It disrupts the grammar, takes a position outside it, uses that position to make the grammar visible as a grammar rather than as nature. Its source material is the still-live dataset: visual work circulating in culture before it has been named, categorised, absorbed, or documented at scale — street aesthetics, subcultural visual languages, formal experiments still being made in marginal publication before they have travelled from avant-garde practice to design blog to corporate application.

By the time a visual style is well-represented in a training corpus, it is already in the absorption phase. The model operates, by structural necessity, on what culture has already half-processed. It can accelerate the distribution of a style that has been named; it cannot be at the frontier of a style that has not yet been legible enough to document. Walter Benjamin's argument in The Work of Art in the Age of Mechanical Reproduction (1935) applies directly, with a sharper edge: when a visual grammar becomes reproducible, it is, by that very fact, no longer at the frontier. Reproduction is the mechanism of legibility, and legibility is the end of the avant-garde function for that gesture.

What AI adds to this dynamic is speed. A visual language that once required years to travel from an avant-garde publication to a mainstream corporate application — the Bauhaus grid entering corporate branding, punk typography entering magazine design, brutalist web aesthetics entering fintech landing pages — can now make that journey in months. The absorption phase compresses; the exhaustion phase arrives faster; the demand for the next genuine gesture intensifies sooner. Doshi and Hauser document exactly this. When the machinery of absorption operates at scale and speed, the scarcity value of the original gesture — the thing that was genuinely new before the cycle began — increases. The machine is doing the derivative work faster. It is still derivative work. Someone still has to produce the thing it will become derivative of.

This is what AI is pointing at when it advises developing taste. The advice indicates the right territory — the territory where AI cannot operate — but uses the wrong word. What cannot be reproduced is not taste in the sense of style management, which is largely already automated. What cannot be reproduced is the vanguard capacity, and behind that, the imaginative capacity that allows the vanguard gesture to be held in the mind before it is made.

IV. Imagination as a technical requirement

Imagination, in AI-native design work, is not a creative virtue or a personality trait. It is a technical requirement of a specific interface paradigm — the capacity to hold a precise internal image of an intended outcome before committing to the command that will produce it.

The etymology is worth holding: imago, Latin for image, for likeness, for the representation of a thing in the mind. In Roman usage the word referred specifically to the wax portrait masks kept by aristocratic families to represent their ancestors in funeral processions — an image held in place of a presence, a substitute for a body no longer available for direct experience. The implication is precise: imagination is not free-ranging vision but the specific capacity to hold something concrete in the mind in the absence of the external referent. Different from creativity in the general sense, from curiosity, from aesthetic sensitivity. It is the ability to see, with enough precision to act on, something that does not yet exist in the world.

This requirement was always present in design work, but it was masked. The real-time feedback architecture of graphical tools was an extraordinarily efficient compensation mechanism for partial vision. Drag, observe, adjust, release. Each cycle cost fractions of a second and nothing else. The hand and the eye did the work that a less practised visual imagination could not do, and the result was indistinguishable from the result of a designer who had seen the composition clearly before beginning. The tool was, in a meaningful sense, a prosthetic for internal precision — and the prosthetic was so good that an entire generation of designers built distinguished careers without ever needing to develop the precision the tool was substituting for.

When the prosthetic is removed — when the feedback loop extends from milliseconds to seconds, when each iteration carries a cost, when commitment to direction must precede evidence for it — the part of the designer's capability that was built on real-time navigation is suddenly not available. What remains is the internal model, and the question is how complete and precise that model is. For designers who built their practice on graphical tools, the first encounter with command-line AI often feels like losing a limb: not because the capability is weaker, but because the capability that remains is exposed as a distinct thing from the capability that was compensating for it.

This applies at every level. The senior designer with twenty years of Figma experience may have built extraordinarily refined judgment through thousands of hours of real-time navigation without ever having needed to hold a complete composition in the mind before the first mark is made. Whether that capacity can be built in mid-career — whether it requires a different relationship to drawing, to mental rehearsal, to the cultural observation that precedes the design act — is a question the profession is only beginning to take seriously.

Figma's canvas-native agent restores some of the feedback loop, and for the navigation of convergent problems this is a genuine return of something like the graphical tool's real-time advantage. But the limit matters. The canvas agent accelerates the iteration phase of a process that has already begun. It does not accelerate the phase before iteration: the observation, the cultural reading, the decision about which direction is worth iterating toward. That phase requires the internal image to exist before the agent is opened. The agent cannot help you decide what you want. It can only help you get there once you know.

For designers who have, or who will build, the capacity to hold a precise internal image before the command, the AI tooling is an accelerant of a very specific kind. The base layer of a project that once consumed weeks arrives in hours. The coverage work that once required careful verification is generated and checked. The responsive variants, the accessibility pass, the component library application, the iteration set: all of this is cheaper and faster, which means the designer's time is available for the work the machine cannot retrieve from its training data. What that work is — and why the corpus structurally cannot reach it — is the final move of the argument.

V. Synthetic imagination drifts from reality

Everything so far rests on a deeper premise that should now be made explicit.

Generative systems do not encounter the world directly. They encounter representations that have already survived social filtering, amplification, documentation, circulation, and absorption. The result is not imagination in the human sense but recursive cultural recombination.

The model does not observe mortality, shame, erotic tension, class aspiration, ageing, illness, humiliation, grief, heat, exhaustion, risk, or desire directly. It inherits their symbolic residue after culture has already processed them into legible form. The image of grief in the training data is not grief; it is the visual language grief has already been translated into — by photographers, illustrators, film directors, advertising creatives, and the algorithms that selected which of their work travelled. By the time grief reaches the model, it is a pattern of compositional cues, palette choices, and gestural conventions. Real grief is somewhere else, and the model has no apparatus to reach it.

This creates drift. Not because the machine becomes detached from culture, but because it becomes increasingly saturated with culture after consolidation. Each generation absorbs the previous consensus and produces work shaped by that consensus, which becomes input for the next absorption cycle. The frontier — the place where new symbolic material is actually being made — sits outside this loop entirely. It sits in subcultural aesthetics, emerging emotional grammars, marginal visual experiments, unstable symbolic forms, and lived contradictions still circulating below visibility thresholds. Wild data.

Social algorithms already industrialised a version of this process, and the distinction is crucial. They amplify emerging signals. They do not surface truly unknown ones. They reward acceleration of legibility, not the illegible frontier itself. A signal becomes algorithmically visible at the exact moment it has begun to be readable to enough people that engagement metrics rise — which is to say, at the moment its raw frontier value has already started to decline. The platforms that look like vanguard detection systems are in fact absorption accelerators. They compress the distance between emergence and consensus, narrowing the window in which a gesture can remain genuinely new.

The frontier, as a result, keeps narrowing. The machine absorbs faster while culture produces genuinely new symbolic material more slowly — partly because the conditions for slow, illegible, first-order experience are themselves being eroded by the same systems doing the absorbing. The compression produces an aesthetic exhaustion that designers can already feel in the corpus: the sameness across AI-generated brand identities, the recognisable signature of synthetic illustration, the way that a sufficiently AI-saturated visual environment starts to refer only to itself.

Artists matter here not as mystical figures but as something more structural. They remain, when they remain, attached to first-order experience before it becomes representational consensus. They observe what is happening in a body, in a neighbourhood, in a generation, in a material, before that observation has acquired its eventual symbolic form. The outsider, the subcultural participant, the emotionally exposed person, the person living inside a contradiction that has not yet been named — these are the inputs the model cannot generate from its existing corpus. They are not romantic figures of authenticity. They are the first-order input layer of future machine imagination.

This is where the economic argument finally lands. The machine's absorption rate is accelerating; the supply of genuinely new symbolic material is not. As the gap widens, the value of work that is actually upstream of the corpus — work made from observation that has not yet been processed into legibility — increases. Not because the work is rare in some artisanal sense, but because the structural position from which it is produced is rare and getting rarer. The designer who can occupy that position, who is willing to attend to what is not yet legible, who can hold an internal image of something the corpus does not contain, is performing a function the machine cannot perform from its training data and will not be able to perform in the next training run either, because the next training run will only contain what the cycle has already absorbed.

The command line never went away. What changed is the price of not knowing what you want to say before you open it.

The capacity to know — to hold a precise internal image grounded in observation that the corpus does not yet contain — is what imagination, in the technical and not romantic sense, now refers to. It is built by being in the world in a way that is genuinely prior to the corpus: attending to what is not yet legible, observing what circulates before absorption, refusing the comfort of conventions already documented at scale. The grief cycle's failure mode was motion without resolution. The acceptance stage, when it comes, does not look like comfort with AI-augmented workflow. It looks like the redirection of attention freed by the transfer of convergent work toward the divergent work the machine cannot do — and toward the cultural exposure that the divergent work depends on.

The designers who will matter in the next five years are not the ones with the most fluent AI workflow. They are the ones whose vision is grounded somewhere the corpus has not yet reached.

References

Benjamin, W. (1935/2008), The Work of Art in the Age of Mechanical Reproduction, trans. J.A. Underwood, Penguin. The argument that mechanical reproduction strips the artwork of its aura; updated here to apply to style rather than to individual works, and to AI reproduction rather than lithographic or photographic copying.

Doshi, A.R. and Hauser, O.P. (2024), empirical documentation of AI-driven homogenisation in creative output. The narrowing of variation in AI-assisted work maps onto the absorption–exhaustion cycle running at an accelerated rate.

Eco, U. (1962/1989), Opera Aperta / The Open Work, Harvard University Press. The structure of bad taste, kitsch as prefabrication of effect, and the relationship between aesthetic convention and interpretive demand.

Eco, U. (1964), Apocalittici e Integrati, Bompiani. The dialectic between critics and enthusiasts of mass culture; the mechanism by which formal novelty is absorbed into consumable convention, and the cycle of disruption, absorption, exhaustion, and disruption that structures modern cultural production.

Figma (2026), "The Figma Design Agent is Here," Figma Blog, May 20. The design intent and architecture of the canvas-native AI agent, including the framing of the speed-versus-precision false choice.

Kessler, D. (2019), Finding Meaning: The Sixth Stage of Grief, Scribner. The argument that healing occurs not when grief gets smaller but when life gets larger — and that the loss itself can become the material of what is built next.

Kübler-Ross, E. (1969), On Death and Dying, Macmillan. The original five-stage framework; Kübler-Ross was consistent that the stages describe a terrain, not a sequence.