As New York Fashion Week descends, editors, buyers and industry watchers flock to the Big Apple amid the usual assortment of excitement and intrigue. Which of the world’s premier designers will impress as they unveil their latest lines and sartorial creations? It is a sign of the times that even this glamorous event is not immune from questions about artificial intelligence: What happens when the synthetic enters an industry steeped in taste and tradition? Can a machine trained on the accumulated imagery of human culture help designers imagine the strange, alluring and occasionally irrational artifacts that constitute high fashion? Or, to borrow from a certain Miranda Priestly, will the outputs of generative systems forever remain an uninspiring “pile of stuff”? The answer may surprise you. 

Traditional A.I. and fashion 

For the past decade, artificial intelligence functioned as a silent partner in the fashion industry, operating largely behind-the-scenes in inventory management, price optimization and predictive analytics. 

Until recently, it was inconceivable that A.I. would intrude on the creative process in an industry defined by the singular vision of its designers. Today, luxury brands are leveraging machine intelligence to amplify the very craftsmanship that distinguishes them from their peers. The role of generative A.I. in this process, and the extent to which it can meaningfully contribute to creativity in fashion, presents a complicated picture.  

Generative A.I. in the design workflow

To understand where generative A.I. enters the creative process in fashion design, it is helpful to consider the basic journey through which an idea becomes a finished garment:

• Archival and trend research: Designers traditionally spend weeks or months assembling references from fashion archives, museums, photography, art and previous collections to establish the visual language of a new collection.
• Concept and silhouette exploration: Those references begin taking physical form through sketching, mood boards and experiments with proportion, shape and drape on dress forms.
• Material and colorway prototyping: Promising concepts are explored through fabric swatches, custom dyeing, material combinations, sample cutting and variations in color and texture.
• Technical design hand-off: Once a design has matured, its creative vision must be translated into technical flats, patterns, measurements and specifications that patternmakers and ateliers can turn into an actual garment.

Within this workflow, the term “using A.I.” can describe dramatically different levels of creative agency. For example, a designer querying a public model for 20 imaginative jacket concepts is doing something quite different from a team that trains a proprietary system on an internal archive and then channels selected outputs into a development pipeline. For the purposes of this analysis, our interest lies in those areas where generative A.I. participates in the creative process itself; instances where the technology can augment, provoke and influence the imagination of the designer. 

In this context, a dividing line emerges between where generative A.I. is most useful and where it’s not. On one side lies the “divergent frontend” of the workflow, where A.I. excels at surfacing references, proposing unexpected combinations and creating visual hypotheses. On the other side, designers leverage their experience in areas where large language models (LLMs) struggle: deciding which ideas are worthwhile, preserving the house’s aesthetic identity and translating images into objects that can be patterned, sampled and produced.

At the divergent frontend, generative A.I. can compress timelines from weeks to hours by allowing designers to explore variations of shape, material and color in rapid succession. For example, a designer might use an LLM to articulate themes and produce atmospheres, objects, environments or speculative materials that would be laborious to source through search alone. In this way, a creative director can quickly evaluate a wide array of visual directions before moving ahead with physical sampling.

In addition, generative systems can envision impossible or unavailable materials—metallic foams, transparent knits, liquid-looking leather—before a studio knows whether they can be manufactured. Such computational creativity can be particularly valuable as fashion innovation often begins with visual contradictions that materials research must somehow make possible. A sprinkle of the impractical can, in other words, widen the creative aperture. 

The more intriguing question, however, is what happens when generative A.I. moves beyond visualization and enters the creative sanctum of fashion itself: the very act of design.  

Generative A.I. and fashion design 

High fashion has always depended on divergence before convergence: the accumulation of sketches, materials and competing ideas before a coherent vision comes into focus. In this phase of the workflow, generative A.I. can be used as an engine for divergent thinking, producing an array of possibilities that would be impractical to render individually by hand. Here, the workflow can take on a curious inversion: early in the process, a designer will give the system some creative latitude, as vagueness can generate surprise, which can in turn spark creative insight. Once an idea has emotional and formal coherence, the designer can apply the opposite pressure: preserving a style or motif while instructing the system to produce incremental variations.  

What is notable in this context is the way generative A.I. can change the psychology of the design task. Whereas traditional sketching gives form to a designer’s thought process one mark at a time, generative systems externalize numerous possibilities at once. Instead of asking “Can I draw what I mean?” a designer armed with A.I. tools might ask “Which of these outputs contains something I mean—or something I didn’t know I meant but want to use?” The workflow can take on an iterative quality in which initial outputs become designs through repeated human selection, rejection and re-prompting. 

A powerful example comes from Collina Strada’s Spring 2024 collection, where designer Hillary Taymour and her team fed the label’s previous work into an A.I. system and spent seven weeks iterating on the results. The model produced exaggerated draping, asymmetries and unusual combinations that felt familiar yet strangely new. Most intriguing were the A.I.’s mistakes: unsupported ruffles, imaginary draping techniques and other visual impossibilities that seemed physically implausible, compelling Taymour’s team to determine how such forms might actually be constructed.  

In this example, artificial intelligence shifted the creative pressure upstream: it generated forms that might not have arisen through conventional sketching and then forced the atelier to devise the technical means to realize them. Such a dynamic can be a powerful model for high fashion as the “error” itself becomes creative fodder, provided the team has enough expertise to distinguish the “inspiringly impractical” from meaningless noise. 

The limitations—going from “bits to atoms”

For all its creative potential, A.I. encounters formidable limitations when the artifacts it generates collide with the physical and cultural realities of fashion—when moving from “bits to atoms,” in the words of artist and designer Lulu Li. 

LLMs can imagine a surface; they cannot certify a textile. Systems like Midjourney and DALL·E will often invent closures that do not close, folds with no structural cause, impossible seam intersections and fabrics that behave inconsistently across the same image. And while an extraordinarily skilled team like Hillary Taymour’s can translate A.I.’s flights of fancy into workable designs, experimentation with generative tools can also be frustratingly counterproductive.    

A second concern is authorship itself. If designers from different houses leverage the same foundational models, there is the risk of “aesthetic convergence,” with ostensibly distinctive brands sampling from the same underlying universe of patterns and imagery. The problem becomes thornier still in the realm of intellectual property: under current U.S. law, prompting an A.I. system alone does not constitute human authorship, leaving purely-A.I. designs without the copyright protection the fashion industry has traditionally relied on. 

And yet such practical concerns are secondary to perhaps the thorniest aspect of leveraging generative A.I. in high fashion: the peculiar psychology of luxury and its aura of rarity. 

The paradox of high fashion 

In mass fashion, products are judged primarily on style, fit and price. Not so for luxury attire, where authorship and process are intrinsic to the product itself, and garments are seen as objects of cultural significance whose desirability depends on provenance, craft, scarcity, continuity and the perceived intent of the designer. A fine jacket, for example, is not simply a possession to showcase on lavish occasions; it carries the story of the label, the authority of a creative director, the labor of an atelier and, most crucially, the aura of rarity. 

In this context, generative A.I. entails a fascinating paradox: while the technology can increase the speed and volume of creative output, it can simultaneously undercut the perceived scarcity of creative labor on which luxury depends. 

To better understand this tension, I spoke with Alexis Lopez Angeles, a fashion designer for Balenciaga. Angeles, who made it clear he was speaking for himself and not his employer, studied fashion at Central Saint Martins in London before completing graduate work at the prestigious Institut Français de la Mode (IFM) in Paris.

“There’s a level of irrationality in high fashion, you know? People don’t buy luxury clothing because they need it. For example, [one of the houses] was doing really big shoes. I own quite a few pairs and they’re hard to walk in. They’re not nice to walk in, they’re not comfortable, but still, they look interesting so people put up with them.”

This irrationality, Angeles says, occasionally gives A.I. an almost surreptitious role when he leverages the technology in his design work: “For A.I., this is all a bit confusing…it can give you the wrong placements, but that can sometimes be nice. I’ve had a couple of combinations where I was like, ‘Oh, this is not what I wanted, but this is interesting, I can build on this.’” 

Artificial Intelligence, he said, can occasionally trigger inspiration in the human designer precisely because its inability to grasp the puzzling sensibilities of consumers can generate surprise. His most penetrating comments, however, were reserved for the psychology of high fashion and the empathy that successful designers bring to the art form. 

“There is one creative director I really like, and when you look at his work you can tell he was relatively insecure about his body. So, he created these silhouettes that were about covering the body in different proportions. It worked because he knew this insecurity also existed in other people,” he said.

“High fashion appeals to a deep level of the human psyche in a way that differs from a product like a phone. If you want to aspire to something creative, you need to understand this psyche and what makes humans tick. Until we have the capacity to bring some sort of subjectivity to A.I., I don’t think we’ll see true creativity [in high fashion].”

LLMs can’t “jump”

Angeles’ observations about the creative limits of A.I. in luxury fashion align with those of prominent technologists. In his recent paper, LLMs can’t jump, DeepMind researcher Tom Zahavy argued that large language models lack a crucial ingredient when it comes to creativity. Although such systems can master induction (finding patterns) and deduction (logical reasoning), they cannot perform the abductive “jump” required to invent entirely new works or theories. 

Zahavy points to Einstein’s Theory of Relativity, which was formulated long before experimental evidence existed, making it impossible to deduce from pattern recognition or logical inference. Rather, he argues, Einstein’s famous thought experiment—imagining an observer falling freely and realizing the equivalence between gravity and acceleration—was fundamentally an act of abduction. That is, history’s most famous scientist mentally simulated a physical situation, experienced its consequences, and generated an entirely new axiom (the equivalence principle) that gave rise to his theory.

The next wave of innovation, Zahavy insists, will come not from increasingly large LLMs, but from interactive, physically grounded models that understand the world. This means closing the loop between sensory simulation and formal reasoning—the missing link required for A.I. to perform Einstein’s abductive “jump.”

The connection between the Theory of Relativity and the latest design from Balenciaga might seem tenuous. But when it comes to high fashion, generative A.I. would surely benefit from a similar “jump” in empathic intelligence and an understanding of the deep, subtle psychology individuals navigate when they wear luxury attire. 

The future—The Devil Wears Prada 3

That artificial intelligence has begun to transform high-fashion design is indisputable, yet its most pronounced effect may lie not in automating creativity but in expanding the design space itself. Generative systems can accelerate early ideation and produce divergent ideas that spark human creativity, but tapping into the needs of luxury consumers requires an empathy that, at present, appears beyond machines. 

To that end, the challenge ahead is no longer “Can a machine draw a beautiful jacket?” but rather “Can it understand the creature who wants it?”

Given the relentless progress of A.I., an assured prognostication on this question would be foolish (to wit, OpenAI’s dramatic announcement that one of its internal models had solved 10 significant problems in math and computer science, a feat that would have been unthinkable only two years prior). And yet one can imagine the following dialogue from The Devil Wears Prada 3, set in 2030:

MIRANDA PRIESTLY (cold and supremely irritated):

“Where are those sketches? I asked for them over an hour ago.”

A trepidatious voice emanates from her computer. It belongs to ChatGPT 10. 

“I’m sorry, Miranda. You requested a new theme only seven minutes ago, and I– “

MIRANDA (forcefully):

“Show me them now.”

ChatGPT renders the entire line on the screen in the new creative direction.  

MIRANDA studies the new collection for three minutes and then whispers:

“That’s all.”

Whether she purses her lips or smiles faintly just before remains to be seen…

Almost a year ago, on the way to soccer practice, my six-year-old asked a question that felt both innocent and quietly inevitable: “Dad, where does ChatGPT live?” Like many children encountering generative A.I. for the first time, he was trying to place it in the world by giving it a location, a biography and a set of boundaries. That instinct matters. Long before children understand how A.I. systems work, they relate to them socially: asking questions, seeking explanations and attributing knowledge, agency and even personality to these tools that respond fluently and with confidence. 

For parents, educators and policymakers, the question is not whether children will interact with generative A.I.—they already are—but rather how those interactions shape formative minds when reasoning, writing and judgment are still being formed. 

My concerns had been simmering long before my son’s query brought them to a boil: What effect would these new tools have on the mind of an impressionable six-year-old? What changes, mundane and grand, would they give rise to in a world he was only beginning to understand? And, finally, what responsibilities did I have as his father to protect and guide him through the very technology that I helped advance in my role as the CEO of an A.I. company?

Parental unease 

The unease around generative A.I.—and its impact on children—is not unique to technology founders or practitioners. Parents, educators and policymakers have all converged around shared concerns about what children should be taught about these systems and what limits, if any, should govern their use. 

Such concerns have become more pronounced as major platforms like OpenAI have moved generative A.I. closer to everyday use. Over the past year, leading models have gained stronger memory, conversational richness and enhanced reasoning, which increase utility for adults while deepening attachment risks for children. 

Every generation, it must be said, laments the negative effects of emerging technologies on younger minds. Radio, calculators, television and even the internet were all once framed as problematic novelties before becoming everyday tools. Generative A.I. differs in a critical respect insofar as it can supplant the act of thinking itself. For children still learning how to reason, write and wrestle with ideas, the concern is clear: what happens when cognitive effort is outsourced before it has been fully developed?

Before grappling with this topic, it is useful to begin with a broader and better-documented question: what does existing evidence tell us about how generative A.I. is already shaping adult cognition? 

Generative A.I. and adult cognition 

The effects of generative A.I. on professional knowledge workers have been thoroughly examined, including a widely cited study by Microsoft Research. Among its central findings is that while generative tools can increase efficiency on particular tasks, their unchecked use can reduce critical thinking. 

The authors describe a phenomenon they term mechanized convergence: users relying on generative systems tend to produce narrower, less diverse outputs than those working independently. A similar pattern was reported in Nature, where researchers found that while A.I. tools expanded the scale of scientific work, they often contracted focus and originality. 

According to the Microsoft study, automating everyday tasks “deprives the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature”. Each time a task—drafting an email, summarizing an article, polishing a LinkedIn post—is outsourced, a small but cumulative cost accrues: the loss of cognitive practice. 

But isn’t the cost worth it, given the productivity benefits? As the refrain goes: jobs will not be replaced by A.I., but by people who can wield A.I. effectively. If this is the case, why shouldn’t such tools be used with impunity? What this framing overlooks  are the cognitive demands required by meaningful human-machine collaboration: strong writing skills to craft precise prompts, and critical thinking abilities to evaluate outputs, identify flaws and iterate thoughtfully. Ironically, these are precisely the skills that erode when generative tools are used without restraint. 

For adults, this paradox raises a key question: how can generative A.I. be used to preserve and amplify critical thinking rather than diminish it? For children, who are only beginning to develop these capacities, the stakes are even higher.  

The use of generative A.I. by young children 

For the purposes of this analysis, young children are defined as those between the ages of four and 12. Precise data on generative A.I. use for this cohort remains limited, in part because many platforms restrict formal access for younger users and usage often occurs indirectly through parents, siblings or educational tools. 

Still, recent research offers a useful window into emerging patterns. A 2025 report by the Alan Turing Institute found that 16 percent of children aged eight to 12 in the U.K. reported using generative A.I. tools daily, with nearly half using them at least weekly. A U.S. study by Common Sense Media reported lower direct use among younger children—six percent of those aged five to eight—but significantly higher exposure in educational contexts, where nearly one-third encountered A.I.-enabled learning tools. 

Taken together, these findings suggest that while daily chatbot use remains relatively uncommon in early childhood, interaction with generative A.I. increases markedly in late primary and pre-teen years. As A.I. tools become embedded across educational systems, productivity platforms and consumer devices, this exposure is likely to grow structurally rather than episodically. 

In practice, the use of generative A.I. in early childhood tends to cluster around four overlapping activities: information seeking, creative play, conversational interaction and homework support, including summarization, drafting and step-by-step assistance.

While many parents report increased engagement and curiosity, concern is widespread. In the U.K. study, more than two-thirds of parents expressed unease, citing risks such as misinformation, inappropriate content and overreliance on automated assistance during formative stages of development. 

Generative A.I. and child cognition 

As with adults, the evidence suggests that generative A.I. can either support or undermine learning depending on how it is used. Across disciplines and age groups, a growing body of research points to a consistent conclusion: A.I. tools enhance learning when they prompt explanation, evaluation and reflection, but undermine it when they replace thinking, practice and persistence. 

The mechanisms behind the benefits and risks of the technology are instructive:

• Cognitive offloading: children may outsource memory, planning or reasoning to the model, reducing opportunities to build and bolster foundational skills.
• Automation bias: students may accept confident-sounding answers without verification, especially when tools are anthropomorphized.
• Scaffolding and formative feedback: when prompts require students to justify, compare or explain, generative A.I. can support higher-order cognition and deepen comprehension.
• Reduced productive struggle: when models supply completed solutions, children lose opportunities to wrestle with problems, make errors and build fluency through effort.

The need for scaffolding and developmentally appropriate tools echoes the findings from older student populations. A field experiment in high school mathematics, for example, found that A.I. tutors lacking strong pedagogical guardrails reduced learning outcomes despite improving short-term productivity. An MIT study on adult essay writing similarly reported that 83 percent of participants who used generative A.I. could not recall a single quote from their own essay, whereas 89 percent of participants who did not use A.I. assistance were able to do so.

Such findings illustrate the pronounced risks of generative tools in early childhood, when critical thinking is built through repeated cycles of asking questions, checking evidence, comparing alternatives, explaining reasoning and reflecting on errors. When A.I. is used as a shortcut—when it becomes a substitute for thinking rather than a tool to provoke it—it can crowd out the very processes that learning depends on. 

For this reason, the consistent guidance is that generative A.I. should be positioned not as a solution generator but as a reasoning catalyst. That is, a tool that generates options, counterexamples or partial explanations for children to interrogate. This reframe is especially important in elementary education, where reading, writing, numeracy and executive function are still developing. 

Policy implications 

Not surprisingly, these concerns have prompted educational and governmental bodies to formalize clear policies on the use of generative A.I. in learning environments. In recent years, guidance has emerged from authorities in England, California, Massachusetts, the government of New South Wales in Australia, as well as multilateral organizations such as UNICEF and UNESCO. While these bodies differ in emphasis, their recommendations converge on a shared concern: generative A.I. must be used in a manner that protects formative cognition while supporting meaningful learning.

In concert, these frameworks articulate some key principles that educators and parents would do well to keep in mind:

• Use bounded tools for younger children: prioritize managed tools with strong filtering, logging, and privacy controls, and avoid open, general-purpose chatbots with weak content moderation or terms that restrict under-13 use.
• Design prompts that force reasoning: require students to explain outputs, identify errors, compare sources and justify conclusions.
• Prioritize process artifacts: emphasize drafts, planning notes, checklists and reflections (e.g., “What did you change after checking?”) to ensure visible thinking remains central to the learning process.
• Teach verification literacy early: introduce age-appropriate instruction on hallucinations, confidence versus correctness, source checking and bias; establish norms such as “A.I. is a starting point, not an authority.”
• Communicate with families: make transparent which tools are approved, the data they collect and the expectations that apply at home (e.g., shared use in common areas, time limits and prohibitions on private, unsupervised chats).
• Protect time for non-digital learning: ensure that reading, handwriting, mental math and unstructured play remain substantial and robust, such that generative A.I. supplements rather than displaces foundational skill-building.

These policy responses reflect a growing recognition within educational communities: the question is no longer whether generative A.I. can assist with learning, but rather how it can be integrated without eroding the conditions that learning depends upon.

Out-of-the-box thinking 

Much of the current discourse has been reactive, emphasizing mitigation, guardrails, and the assimilation of generative A.I. into existing pedagogical structures. While valid, a parallel question deserves attention: how might education be reimagined with generative A.I., rather than merely defended against it? 

To this end, a more forward-looking perspective is offered in a recent whitepaper, AI and the Future of Pedagogy. While expanding on the cautionary points above, it also explores how generative A.I. might enable new forms of learning that were previously impractical. A single example is instructive. 

Consider the topic of student testing. Instead of relying on static, fact-based assessments, the whitepaper imagines an adaptive, Socratic evaluation environment powered by generative models. In this paradigm, the assessment becomes a guided conversation that probes student comprehension in a variety of directions, and shifts the underlying question from “What do you know?” to “How do you think?” Competence is measured not by recall, but by the ability to apply concepts in novel contexts, articulate reasoning and connect disparate ideas; capacities that align more closely with the cognitive skills education ultimately seeks to cultivate.

An educational system that fuses thoughtful experimentation with the wisdom of its practitioners is a credible cause for optimism. 

A look to the future 

Futurist Rahim Hirji observes that “by 2030, the question will no longer be whether automation replaces jobs…[but] what humans choose to become once intelligence is abundant.” In this framing, the long-term stakes of technology will depend less on the mastery of specific systems and more on qualities that remain eminently human: curiosity, agency, empathy, integrity and the capacity to ask better questions rather than simply retrieve better answers.

In their fourth Economic Index Report, Anthropic, creator of the leading model Claude, developed a metric called “human education years,” an estimate of the number of years of formal schooling an individual requires to understand a user’s prompts and the model’s responses. 

A key finding is that the sophistication of the model’s answer closely mirrors the complexity and nuance of the user’s prompt. In other words, A.I. does not nullify the fundamentals; it accentuates them. Core skills centered on reading, clear writing and structured reasoning are not redundant in the age of A.I., but prerequisites for extracting its value. These capacities must be cultivated deliberately, particularly in childhood, as generative systems continue to grow in capability and influence. 

Returning to my son’s original inquiry, how might a parent respond to their child’s question of “Where does ChatGPT live”? One might encourage them to ask the model that very question and then probe it with relentless curiosity—how, when and why. But, in the spirit of this analysis, the key lies in the follow-up thereafter, one that conscientious parents of generations past knew well: lively discussion, tender mentorship and informed debate. 

The responsibility now falls to adults to model and foster a point that will become increasingly important in this age of technical adolescence: A.I. doesn’t replace thinking—it rewards it.