By Dave Fuller, Managing and Technical Director, Accent
The question everyone asks first is the wrong one
"Can AI be creative?" is a question that generates a great deal of heat and almost no useful light. It collapses two very different things into one word.
If creativity means the production of novel, competent, well-formed artefacts, then yes, obviously. AI does this at industrial scale, at near-zero marginal cost, in seconds. That argument is over and it was over quickly.
If creativity means the origination of intent, the decision that this thing is worth making, for these people, for this reason, then the picture is entirely different. And it is that second definition that carries all the commercial value.
The useful question is not whether machines can be creative. It is: where in the creative process does human judgement still change the outcome, and where has it stopped mattering?
That question has a defensible answer. It also has direct, near-term consequences for anyone running or buying web design work.
What the publishing industry just learned the hard way
In late 2025, Hachette, one of the largest publishers in the world, acquired a self-published horror novel that had already found an audience. This is increasingly normal: traditional publishers now treat self-publishing as a proving ground, picking up titles that have demonstrated readership rather than guessing at it.
The book was subsequently pulled and pulped after the publisher concluded it had been substantially written or amended by AI. The author's account was that she had passed a draft to someone in her writing group who used ChatGPT to edit it, and that she had failed to do a final careful pass before publishing.
Several things about this episode are worth dwelling on, because they generalise far beyond books.
The detection didn't come from detection software
This is the detail most commentary misses. The suspicion originated on a Reddit thread, from readers. It was confirmed by editors and agents who, as discussed on The Rest Is Entertainment, overwhelmingly do not use AI detection tools. They use their judgement, built over decades of reading manuscripts all day, every day.
The tells they describe are structural rather than lexical. Every noun carries an adjective. Every action attracts a simile. Relentless tricolons. Sentences that sound profound and mean nothing on second reading. And crucially: an emotional flatness across the whole, an inability to modulate.
The point made on the podcast is precise and worth quoting the substance of: every one of these devices is something human writers use. The difference is that humans do not use all of them on every page. A human writer reaches the end of a chapter and asks what the reader is feeling, whether the section was too dense, what move would surprise them next. The model asks what word would normally come next.
That is the actual technical boundary, and it is not going away with the next model release. Next-token prediction optimises for the plausible continuation. Creative work frequently requires the implausible one. These are different objectives, and no amount of scaling reconciles them, because the thing being scaled is the very property that produces the flatness.
Detection tools are largely theatre
Someone ran the manuscript through a scanner and produced a figure of "78% AI". Defenders of the author responded, correctly, that running Frankenstein through the same tools returns 100% AI.
Both facts are true and both are useless. The scanners produce a number with no reliable relationship to reality, and much of that market exists to convert anxiety into subscriptions. Anyone building an AI policy around automated detection is building on sand.
The accusation is now a weapon
A debut author had their first book accused of an AI-generated cover in its launch week. The accusation was false, and they could prove it only because the illustrator happened to have recorded time-lapse footage of herself drawing it.
Sit with the implications. That author was saved by an accident of process. Anyone without that footage would have had no defence, because you cannot prove a negative about your own creative process after the fact.
The prediction that follows is that provenance becomes infrastructure. Not a nice-to-have, not a marketing badge, but a documented, contractual part of how creative work is commissioned and delivered. Writers will need the equivalent of the illustrator's time-lapse. Publishers will need to sign to say a brief was human-originated. And it will be established practice within a few years, because the alternative is a market where any competitor can destroy any reputation with a tweet.
There is no copyright in machine-generated ideas
The commercial dimension is sharper still. The genuine strategic interest publishers have in AI is not writing books; it is owning IP without a human originator to pay. Commission the concept from a model, hand it to a jobbing writer, own everything.
Except that under current understanding in both UK and US law, purely machine-generated material attracts no copyright protection. So if a publisher's blockbuster premise came from a model rather than a person, they cannot stop anyone writing the sequel. Or the one after that.
This inverts the incentive completely. The commercial reason to keep humans at the point of origination is not sentiment. It is that human origination is the only thing that produces a defensible asset. Provenance stops being an ethics question and becomes a balance-sheet question.
Where the value is actually moving
With that grounding, the broader structural argument becomes clearer.
1. The collapse of execution as a differentiator
For most of the history of creative work, technical craft was the bottleneck. Being a creator meant mastering a medium: brush technique, camera work, software proficiency, syntax. That mastery took years, and the years were the moat.
AI removes the bottleneck almost entirely. The constraint moves from can you make this to should this exist, and is this the version worth making.
This is not a small adjustment. It reprices the entire field. Skills that took a decade to acquire, and which commanded a premium precisely because they took a decade, are now available on demand. The premium relocates to direction, judgement and taste, which are harder to acquire, harder to teach and considerably harder to fake.
2. The great commoditisation, and the death of the B-minus creative
AI is exceptionally good at "good enough". Stock imagery, serviceable copy, background music, competent-but-unremarkable layout. It raises the floor dramatically. Nothing produced from now on has any excuse for being incompetent.
But raising the floor is not the same as raising the ceiling, and the effect on the middle is brutal. The commercial creative whose value proposition was reliable, professional, on-brief execution at the B-minus tier is competing directly with something that does B-minus instantly and for nothing.
Meanwhile, the premium on genuinely disruptive work goes up, precisely because a model optimised on existing patterns cannot produce work that breaks them. Scarcity moves to the top of the distribution.
The result is a shrinking creative middle class. And this creates a genuine, unsolved structural problem that the industry has not begun to answer honestly.
The junior rungs are being sawn off the ladder. Junior copywriters, storyboard artists, production designers, entry-level front-end developers: these roles existed partly to produce work and partly as apprenticeship. Ten years of doing B-minus work under supervision is how people historically developed the judgement to become creative directors.
If the entry-level work is automated, where does the next generation of senior judgement come from? Nobody has a good answer. The honest position is that the industry is currently consuming a stock of experienced judgement it is no longer replenishing, and the bill for that arrives in roughly a decade.
There is a concrete version of this that plays out in our own studio, and it centres on the most automatable skill in front-end development.
HTML, Tailwind and native CSS are exactly the sort of thing models generate fluently. Ask for a responsive card grid and you get one, immediately, and it will probably work. On the surface this makes learning to write markup and styles look like a waste of a junior's time.
It is the opposite, for a reason that is easy to miss: the AI only succeeds when the person directing it already understands what should come out. Someone who knows how the cascade actually resolves, what the semantic elements are for, when a utility class is the right call and when it is papering over a structural problem, can look at generated output and see immediately that the heading levels skip from h2 to h4, that the div should have been a button, that the layout will collapse at a breakpoint nobody tested, that the class list has grown to forty utilities because the underlying structure is wrong.
Someone without that grounding sees code that renders correctly in the preview and ships it. The failure is invisible until it reaches a screen reader, a real device, or a maintenance pass eighteen months later.
So the tedious version is not obsolete busywork. Writing the markup by hand is how you acquire the ability to evaluate markup you did not write, and that evaluative ability is now the entire job. The generation is free; knowing whether it is right is the thing being paid for.
Any agency thinking seriously about the next ten years should treat this as a strategic problem rather than a hiring one, which in practice means deliberately keeping juniors on work that a model could have done faster.
3. The curation artist
The corollary of removing the execution bottleneck is a genuine democratisation of production. A novelist can make a film. A musician can build a game. The barrier to entry approaches zero.
What follows is an explosion of niche, hyper-specific work that would never have justified its production cost before, and simultaneously an overwhelming sea of undifferentiated noise.
In that environment, the scarce resource is not the ability to produce. It is the ability to choose. The new creative elite are not the people who can draw the straightest line; they are the ones who can ask the most evocative question, recognise the one good output among forty mediocre ones, and know when to throw the whole batch away.
Creativity becomes an act of editing and curation as much as generation. But this comes with an important caveat that gets glossed over in the optimistic version: you cannot curate what you cannot evaluate. Taste is not innate. It is built from having done the work, having failed, having seen what a real solution costs. The curation-first model works brilliantly for people who already have twenty years of craft behind them, and much less well for people who have skipped straight to prompting. This connects directly back to the apprenticeship problem, and it is the same problem wearing a different hat.
We have a first-hand example, and it is slightly embarrassing, which is why it is worth including.
When we wrote up how AI cut roughly a third off our monthly project management, drafting assistance repeatedly invented client project specifics. Not vague filler: plausible, confident, specific detail about systems and decisions that did not exist. Every fabrication read as competent. None of it would have been caught by anyone unfamiliar with the actual projects.
The more instructive failure was subtler. Some of the invented material actively worked against the argument we were making. The piece was about the necessity of human verification, and the drafting process kept producing exactly the kind of fluent, unverified content the article warned about. Fixing it required someone who both knew the client history and held the editorial line, and the fabrications that undermined the narrative were harder to spot than the outright false ones, because they were internally consistent.
This is the practical shape of "you cannot curate what you cannot evaluate". The output was not obviously bad. It was obviously bad only to someone who already knew.
4. The aesthetics of imperfection
The final dynamic is a consumer-side reaction, and it is already visible.
As machine-generated content becomes ubiquitous and uniformly polished, a fatigue sets in. Human creative work has historically been defined partly by its constraints and its flaws: the happy accident, the physical limits of the medium, the grain, the wobble, the emotional weather of the person making it. Optimisation systems are built to eliminate exactly those things, which is why the output can feel sterile, uncanny or hollow, even when technically flawless.
The market response is a premium on the authentically imperfect. The resurgence of vinyl, film photography, letterpress, hand-drawn animation and visible brushwork is not nostalgia. It is a demand signal for evidence of human involvement.
The podcast's framing captures it neatly: the analogy to a National Trust gift shop, where the pot of jam is signed by the person who made it. An artisan newspaper column. A gold standard for human-made.
Expect "100% human-made" to function as a luxury marketing claim, in the way "hand-stitched" or "single origin" does. Expect it to be abused. And expect it to matter most in exactly the categories where trust is the product.
What this means for web design specifically
Web design sits at an unusually exposed intersection: it is simultaneously a craft discipline, a commodity service and a trust-signalling medium. All three of the above dynamics land on it at once.
The template tier is finished, and that is fine
The lower end of web design, the brochure site assembled from a template with stock photography and generic copy, has been under pressure for fifteen years from page builders. AI finishes the job. That work is no longer viable as a business, and defending it is a waste of energy.
What replaces it is not nothing. It is a higher floor: clients who previously received a bad template site now receive a competent generated one. The agencies that survive this are the ones that were never really selling page production.
Production velocity is now table stakes, not a differentiator
AI genuinely does compress parts of the process, but the compression is not where most people assume. The headline saving in our own studio has not come from code generation. It has come from an admin function nobody ever wanted to pay for: requirements gathering.
Discovery on a complex project produces information across an absurd number of channels. In-person meeting recordings. Teams transcripts. Handwritten notes. Email chains. Spreadsheets and Word documents sent as attachments with no summary. Collating that, deduplicating it, reconciling the version where the client said one thing in a meeting and something slightly different by email a week later, then turning it into a single reviewed statement of requirements, is weeks of unglamorous work. Historically it could run to months on a large project.
That now takes days. Multiple streams get transcribed and synthesised into one structured picture, which a human then verifies against the source. Two things follow, and the second matters more than the first.
Nothing gets missed. The quiet constraint mentioned once, forty minutes into a call, no longer falls through the gap between someone's notebook and the follow-up email.
And we report back while the meeting is still fresh in everyone's mind. This is the underrated part. A requirements document returned three months after the workshop asks the client to remember what they meant. Returned in three days, it lands while the reasoning is still live, so corrections are cheap and accurate. Compressing the calendar gap changes the quality of the client's own input, not just our turnaround.
This is documented in more detail in our write-up of the project management workflow, where the same approach reduced monthly project management admin by around 70% over eighteen months.
Two honest caveats. Every synthesised summary is read against the recording before anything is acted on, which costs roughly ten minutes per hour-long call and is non-negotiable. And when every competitor has the same compression, it stops being an advantage and becomes an expectation. Speed gets competed away. The question becomes what you do with the recovered time, and the answer needs to be "more thinking", not "more sites".
Where the defensible value concentrates
Strategy and information architecture. A model can produce a navigation structure. It cannot work out which structure the site actually needs, because that requires understanding who is arriving and what they came to do.
Our work with the Eastern Shotokan Karate Association in Norwich is a useful illustration. The presenting problem was a site that was slow, unreliable and awkward on mobile. Those are real problems and we fixed them: a 95% Google Page Speed score, mobile-first throughout, hosting and broken links sorted.
But the structural issue underneath was that the site had to serve two audiences with almost opposite needs from the same set of pages. A prospective member who has never trained before needs reassurance, instructor faces, and an obvious answer to "when and where do I turn up". An existing karateka needs schedules, grading information and club material, and needs it in two taps because they are checking on a phone before a session. A single navigation optimised for one of those audiences fails the other.
The resolution was to separate them properly: a public-facing route built around discovery and confidence, and a members-only zone for the people already in the club. Instructor profiles with moving portraits were not decoration; putting faces on the site addresses the specific anxiety that stops beginners walking into a martial arts club for the first time.
None of that came from the brief. It came from understanding a karate club, which meant asking why people hesitate before joining one. A model given "redesign this karate site" produces a faster version of the same wrong structure.
Distinctiveness as a business asset. If every competitor in a sector generates their site from models trained on the same corpus of existing sites, every site in that sector converges. Sameness is the default output. A brand that looks meaningfully different becomes commercially valuable because difference is now scarce and expensive. This is the "ceiling" argument applied directly: the premium on genuinely distinctive design goes up, not down.
Conversion and behavioural judgement. Knowing that a form converts better with fewer fields is table stakes. Knowing which fields, for this audience, at this point in a considered purchase, is domain knowledge plus evidence, and it is the thing clients are actually buying.
Accessibility done properly. This is where the gap between plausible output and correct output is widest, and where it carries the most exposure.
Generated markup looks accessible. It has alt attributes, ARIA roles, labelled form fields. It will pass an automated scan. But independent research from Deque, WebAIM and others consistently puts automated coverage at somewhere between a quarter and 40% of WCAG success criteria, depending on how you measure. The remainder covers reading order, focus indicator visibility, contrast in context, whether the alt text actually describes the image, and whether the instructions make sense to a person using them. All of that requires human judgement, testing with assistive technology, and ideally testing with disabled users.
There is a sharper detail that ought to worry anyone leaning on generated markup. The WebAIM Million study has repeatedly found that pages using ARIA have more detected accessibility errors than pages without it, not fewer, because ARIA is so frequently misapplied. A model will happily add role and aria-label attributes to a div that should simply have been a button. The output looks more accessible and is measurably less so. The scan comes back clean, and the page is broken for the people it was supposed to help.
The regulatory direction is one way. The European Accessibility Act took effect in June 2025, and accessibility litigation in the US has continued climbing year on year. A clean automated report is not a defence, and "the tool generated it" is not a defence either. Someone has to be accountable for whether the thing actually works, and that accountability cannot be handed to a system that cannot hold it.
Integrity of the build. Generated code is confidently wrong in ways that pass review if nobody senior is reviewing. The value of a person who can read it and know it is wrong goes up as the volume of generated code goes up.
The provenance question arrives here too
The publishing episode is a preview. Clients will increasingly ask what in their site was machine-generated: for copyright reasons, for brand-integrity reasons, and eventually for regulatory reasons.
Agencies that can answer that question cleanly, with documented process, will be in a materially stronger position than those improvising an answer under pressure. This is worth building now, while it is cheap, rather than later, under duress.
Practically: keep the working files, keep the version history, document where generated material was used and how it was reviewed, and be able to state what was originated by a person. This is not paranoia. It is the same instinct as the illustrator's time-lapse, applied deliberately rather than by luck.
The honest counter-argument
An article of this kind should state the strongest case against its own position.
That case is: models improve, and the flatness described above is a current limitation rather than a permanent one. Each generation is measurably better at holding structure and varying register. Betting a business on "AI can't do taste" may be betting on a receding property.
There is force in this. The response is twofold.
First, the objective function argument stands. The commercial pressure on frontier models is towards reliability, accuracy and consistency, because that is what the enormous markets, logistics, medicine, finance, insurance, actually pay for. The podcast makes this point sharply: creative writing was never the destination, it was training material for a repository of well-formed sentences. There is no meaningful money in radical, weird, disruptive output, and so there is limited commercial pull towards it.
Second, and more fundamentally: even granting a model that could produce genuinely surprising work, someone still has to decide what problem is worth solving, judge whether the output solves it, and take responsibility when it does not. Accountability cannot be delegated to a system that cannot hold it. In client services, that is not a technicality. It is the entire basis of the relationship.
The position, stated plainly
AI has made execution cheap and abundant. It has not made judgement cheap, and it has not made accountability transferable.
The human role is not decoration on an automated process. It is origination, direction, evaluation and responsibility. The spark, and the decision about which sparks are worth following.
For the web design industry, the implications are specific:
- Work whose value was execution speed is being repriced towards zero. Accept it and move.
- Work whose value is strategic judgement, distinctiveness, and accountable delivery is becoming more valuable, because the surrounding noise is increasing.
- Provenance and process documentation are becoming commercial infrastructure, for copyright reasons as much as ethical ones.
- The apprenticeship problem is real, unsolved and worth taking seriously now rather than in 2035.
The organisations that struggle will be those that used AI to produce more of the same, faster. The ones that thrive will use the recovered time to think harder about what should exist at all.
That has always been the job. It has simply stopped being possible to hide behind the craft.
Sources and further reading
- The Rest Is Entertainment, Richard Osman and Marina Hyde, episode covering the Hachette withdrawal and AI in publishing (transcript reviewed for this article).
- US Copyright Office guidance on copyright registration for works containing AI-generated material.
- W3C Web Content Accessibility Guidelines (WCAG 2.2).
- WebAIM Million annual accessibility report, on ARIA misapplication and homepage failure rates.
- Deque and UK Government Digital Service analyses of automated accessibility testing coverage.
- European Accessibility Act, in force June 2025.
A note on how this article was produced
The argument, structure and editorial position of this article originated with Dave Fuller, Managing Director at Accent. Source material included a full transcript of the referenced podcast episode. AI tools were used for research summarisation. All claims were verified and the final text was reviewed and edited by a person, who takes responsibility for it.
Dave Fuller is Managing and Technical Director of Accent Design Group Ltd, the Norwich digital agency that has designed, built, hosted and maintained websites and web-based applications for over 27 years. He holds qualifications in web accessibility and design.