There is a particular kind of fashion moment that resists every model in this series. It is the moment a designer looks at a finished garment — technically correct, well-constructed, on-trend by every measurable signal — and decides it is wrong. Not wrong in any quantifiable way. Not wrong by sell-through prediction or trend-alignment score. Wrong in a way that can only be described as a feeling, and that feeling, more often than the data, is what gets the garment cut from the collection.

This article asks a question the rest of this series has mostly sidestepped: can an algorithm have taste? Not whether AI can generate fashion content, optimise prices, predict demand, or detect defects — all of which it demonstrably can, as the previous thirteen articles in this series have documented in detail. The question is narrower and harder: can a machine make the specific kind of aesthetic judgement that defines what fashion is, as opposed to what fashion does commercially?

The honest answer, examined carefully, is: not yet, and possibly not in the way the question assumes. But the more interesting answer is what taste actually is, why it has proven so resistant to computation, and what that resistance reveals about the limits of treating fashion as an optimisation problem in the first place.

Taste is not the same thing as preference, and the distinction matters enormously for understanding what AI can and cannot model. A preference is a stable, individually-held inclination — I like blue, I dislike pleated trousers — and preferences are exactly what recommendation algorithms, the subject of an earlier article in this series, are extremely good at predicting and serving. Taste is something else: a culturally situated, time-sensitive, often contradictory judgement about what is currently right, made in relation to everything that came immediately before it.

This is why a garment that would have been celebrated as visionary in 2019 can feel exhausted and derivative by 2025 — not because the garment changed, but because taste moved past it. Taste is fundamentally relational and temporal: it is a judgement about a thing’s relationship to a specific cultural moment, not an absolute property of the thing itself. A model trained on historical data is, by construction, looking backward. It can identify what has been considered tasteful. It cannot, by the nature of how machine learning works, tell you what will feel exhausted the moment it ships — because that judgement requires a sense of where culture is moving next, not where it has been.

Designers and creative directors describe this as an intuition that operates faster than analysis — a near-instant sense of rightness or wrongness that they often cannot fully articulate even to themselves, let alone encode as a set of rules a model could learn from. Phoebe Philo, discussing her design process in rare interviews, has described decisions made from a feeling she could justify only after the fact, if at all. This is not mysticism. It is the description of a cognitive process — rapid, holistic, culturally embedded pattern recognition — that current AI architectures are not built to replicate, because that process is not reducible to the kind of labelled training data generative and predictive models require.

Almost every AI application covered elsewhere in this series shares a common structure: define a measurable objective, train a model to optimise toward it, and trust the model’s output because the metric improves. Demand forecasting optimises for prediction accuracy. Dynamic pricing optimises for revenue. Computer vision optimises for defect detection rate. This structure works because these are genuinely optimisation problems with a clear, measurable definition of success.

Aesthetic judgement is structurally different, and applying the same optimisation logic produces a specific and important failure mode. If you train a model to predict which designs will perform best by historical sales or engagement metrics, you have built a system that optimises for what has already worked — which is, definitionally, the opposite of what makes a design feel new. The most commercially successful designs by historical metrics are, by construction, the least likely to be the designs that define the next aesthetic era, because that era has not generated training data yet.

This is the optimisation trap: a model that is genuinely excellent at predicting commercial performance based on historical patterns will, if followed uncritically, systematically steer a brand away from the genuinely novel choices that periodically redefine fashion. Every major aesthetic shift in fashion history — the introduction of androgynous tailoring, the rise of streetwear into luxury, the normalisation of “ugly” footwear silhouettes — would have scored poorly against a model trained on the data that existed immediately before that shift, because the shift had not happened yet for the model to learn from.

Demna at Balenciaga has been explicit in interviews about deliberately pursuing design choices that internal data and conventional wisdom would have flagged as commercially risky — oversized silhouettes, deliberately “ugly” sneakers, garments that polarised opinion sharply on first release. Several of these choices became among the brand’s most commercially successful and culturally influential. A model optimising for predicted commercial success, applied at the moment of the decision, would very plausibly have recommended against each of them.

None of this means AI has no role in aesthetic decision-making — it means the role is different from what is often claimed. As explored in this series’ article on generative AI in the design process, AI’s genuine strength is in expanding the field of options a designer considers, not in selecting which option is correct. This is a meaningful and underrated contribution: a designer with a thousand generated variations to react against has more raw material to exercise taste upon than a designer with ten hand-drawn sketches.

There is also a genuine role for AI as a devil’s advocate against a designer’s own blind spots. A model trained on a brand’s historical output can flag when a new design is simply repeating a pattern the brand has used five times before — surfacing a kind of unconscious repetition that a designer working under deadline pressure might not notice themselves. This is not AI exercising taste; it is AI providing the kind of structured self-awareness that a thoughtful creative collaborator might offer, without claiming to know what should replace the repeated pattern.

AI is also legitimately useful at filtering for technical feasibility before human taste is applied — removing the options that are not physically constructible, that violate brand cost targets, or that fail basic fit logic — so that the human curator’s attention is spent on the genuinely viable creative choices rather than wasted screening out options that should never have reached them. This is a meaningful efficiency gain that does not require claiming the AI has aesthetic judgement, only that it can pre-filter for constraints that are genuinely measurable.

A fair response to everything above is that human taste itself is a product of pattern recognition over lived experience — a designer’s intuition was built by years of looking at fashion, absorbing cultural signals, and developing a model, however informal, of what currently feels right. If a human’s taste is the product of a sophisticated pattern-recognition process running on biological hardware, why assume a sufficiently sophisticated artificial pattern-recognition process couldn’t eventually do the same thing?

This is a genuinely serious argument, and the honest response is that it cannot be fully dismissed on principle. The difference, today, is not necessarily categorical — it may be a difference of degree and data. Human taste is built on multimodal, embodied, socially-situated experience: not just looking at images, but feeling fabric, watching how people move through the world in clothes, absorbing social cues about status and belonging, experiencing the emotional weight of specific cultural moments. Current AI models trained on visual datasets alone are working with a radically impoverished version of the inputs that shape human taste.

It is plausible — though not yet demonstrated — that a future system trained on far richer, more embodied, more socially-contextualised data could begin to approximate something closer to taste. What is empirically true today is that no system has done this. The generative models that produce visually compelling fashion imagery are doing something closer to sophisticated recombination of historical visual patterns than to the forward-looking cultural judgement that defines taste at its sharpest. The gap may narrow. It has not closed.

There is a real and documented commercial risk in fashion brands leaning too heavily on data-driven decision-making at the expense of taste-driven judgement — a risk that is becoming more relevant precisely as AI tools make data-driven design decisions easier and more tempting to default to.

Brands that have allowed quantitative performance data to dominate creative decision-making have, in several well-documented cases, drifted toward a kind of aesthetic averaging: designs that test well, offend no one, and feel increasingly indistinguishable from competitor offerings optimising against the same metrics. Gap Inc.‘s extended period of declining relevance through the 2010s has been attributed in part to design decisions increasingly driven by testing and data rather than a strong, distinct creative point of view — a cautionary case frequently cited in fashion business analysis of what happens when commercial optimisation crowds out aesthetic conviction entirely.

The counter-case is equally instructive. Brunello Cucinelli, one of the most consistently profitable luxury houses, has been notably resistant to data-driven design decision-making, maintaining a design process built almost entirely around the founder’s personal aesthetic conviction. The brand’s financial performance — sustained premium pricing power and consistent growth — suggests that strong, undiluted taste, even when it cannot be justified by predictive data, can be a more durable competitive advantage than aesthetic decisions optimised against historical performance metrics. The lesson is not that data is worthless. It is that taste and data answer different questions, and conflating them produces worse outcomes than keeping them appropriately separate.

The taste question in fashion is a specific instance of a broader question that every creative and judgement-intensive industry is currently working through as AI capability expands: which decisions are genuinely optimisation problems with a measurable right answer, and which are judgement calls where the entire value lies in a perspective that cannot be reduced to a metric?

Fashion is a useful test case precisely because it makes the distinction unusually visible. A garment’s commercial performance is measurable in a way that, say, the quality of a piece of writing or the wisdom of a strategic business decision is not. And yet fashion’s own history demonstrates repeatedly that optimising for measurable commercial performance, applied at the moment of creative decision, would have prevented many of the choices that came to define the industry’s most significant aesthetic shifts. If even the most measurable creative industry resists full optimisation, that should inform how cautiously other industries lean on AI for genuinely judgement-intensive decisions.

The brands and creative leaders navigating this well are not rejecting AI — every other article in this series documents how thoroughly AI has earned its place across fashion’s operations, supply chain, and even significant parts of the design process. They are simply refusing to let the fact that AI can generate options be confused with AI’s ability to judge them. That distinction — between generation and judgement, between pattern and taste — is likely to remain the dividing line between what AI does for fashion and what fashion’s most distinctive voices continue to do for themselves.