Photo: Bekky Bekks / Unsplash
THE FRICTION
Audit
A study in Science Advances gave writers AI assistance and measured what came out. Everyone got better. Everything got more similar.
Individually, the AI-assisted work was rated more creative, better written and more enjoyable — with the biggest gains going to the weakest writers.
Collectively, those same pieces were measurably more like each other than work produced by humans alone. The authors' conclusion: writers were individually better off, but a narrower range of ideas survived.
Doshi & Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28), July 2024That is the whole problem in one sentence. Every team adopting AI is getting a private lift and paying a collective cost — and the cost is the thing they compete on.
This report follows that finding through the evidence: controlled experiments, payroll microdata, and four million pieces of real creative output.
What emerges is consistent and slightly unnerving. AI makes almost every individual better. It makes the group's work more alike. And the thing that decides whether AI lifts you or quietly degrades you turns out to be the one thing nobody knows how to buy.
Photo: Raymond Petrik / UnsplashTHE
CONVERGENCE
Everyone gets better. Everything gets more similar. Both effects have now been measured.
Published in Science Advances, from a controlled experiment: AI-assisted writers produced better work — and work that was more like everyone else's.
The authors' own conclusion: "writers are individually better off, but collectively a narrower scope of novel content is produced."
Doshi & Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28), 12 July 2024. The 26% figure is MIT Sloan Management Review's reading of the same study.Four million artworks. Fifty thousand creators. Same result, in the wild.
Researchers tracked a text-to-image platform over time. Productivity rose 25%. The likelihood of a piece earning a favourite rose 50%. And average content novelty fell — with "a consistent reduction in both peak and average visual novelty."
More output. Better received. Less new.
Zhou & Lee, "Generative Artificial Intelligence, Human Creativity, and Art," PNAS Nexus 3(3), 29 Feb 2024 (4M+ artworks, 50,000+ users)Only a quarter of people think they can spot AI creative. It still costs you when they sense it.
The Nuremberg researchers gave it a name. An identical ad, labelled AI-made, was rated worse than when labelled human-made — worst of all on the emotional dimensions. They call it a trust penalty.
Photo: Carlos Torres / UnsplashWHAT AI
ACTUALLY TAKES
Creative leaders' forecasts and the measured evidence disagree — in a specific, useful way.
Creative work is not waiting for AI to arrive. Measured by real usage, it is already one of the most AI-exposed categories of work there is.
Anthropic mapped a million real conversations against the US task database. Arts, design, entertainment and media accounted for 10.3% of all queries — against just 1.4% of the actual workforce. Creative roles are using AI at roughly seven times the rate their share of employment would predict.
So the question was never whether AI enters creative work. It's already in. The question is which parts it takes.
Anthropic Economic Index, Feb 2025 — ~1 million Claude.ai conversations mapped to O*NET's ~20,000-task databaseTwo of the largest skill-level audits ever run found almost nothing that AI can fully take.
Graphic design is now the 11th fastest-declining job employers forecast for 2025–2030 — a reversal from moderate growth two years earlier. UI and UX design is the 8th fastest growing. Execution contracts; judgment expands.
World Economic Forum, "Future of Jobs Report 2025," Jan 2025Across the same data, augmentation outweighs automation — 57%/43% in early 2025, 52%/45% by late 2025. Only about 4% of jobs used AI for 75%+ of tasks.
Anthropic Economic Index, Feb 2025 and Jan 2026In MIT's experiment, AI cut time 37% and raised quality 45%. Rough-drafting time more than halved — while editing time more than doubled. The work moved to judgment.
Noy & Zhang, MIT, March 2023 (444 professionals including marketers)
Photo: Pawel Czerwinski / UnsplashIn the same MIT experiment, 68% of participants submitted the AI's first output without editing it at all.
Which is how you get the numbers below. The judgment layer is where the value moved — and it's the layer people skip.
36.5% of marketers say hallucinated or inaccurate AI content has already gone live. Only 23% are comfortable using AI output with no human review.
NP Digital, 30 Jan 2026 (600 prompts across 6 LLMs; 565 US marketers)51% of organisations using AI have already had at least one negative consequence; nearly a third report consequences from inaccuracy.
McKinsey, "The State of AI in 2025," 5 Nov 2025 (n=1,993, 105 countries)
Photo: Antony Hyson Seltran / UnsplashTHE JUDGMENT
PREMIUM
Taste sounds like a soft word. It is now the hardest-edged variable in the data.
Three independent studies gave people identical AI access. In all three, the user's judgment decided whether AI helped them or hurt them.
Given the same AI advisor, already-strong performers gained 10–15% in profit. Already-weak performers lost 8%. The tool amplified whatever judgment was already there.
Koning et al., "The Uneven Impact of Generative AI on Entrepreneurial Performance," HBS Working Paper 24-042 (640 entrepreneurs, randomised)Inside AI's competence frontier, consultants were 40%+ higher quality. On a task outside it, AI users were 19 percentage points less likely to get it right than colleagues with no AI at all.
Dell'Acqua et al., "Navigating the Jagged Technological Frontier," HBS/Wharton/MIT/BCG, pre-registered experiment, 758 BCG consultantsAI closed the ideation gap between domain insiders and outsiders almost completely. It did not close the execution gap: outsiders still scored about 13% lower on quality.
Vendraminelli et al., "The GenAI Wall Effect," HBS Working Paper 26-011, Sept 2025 (78 workers, IG Group)How you use the model matters more than whether you have it.
Researchers tested LLMs on ad copy, then measured quality by actual clicks on major social platforms. Used as a sounding board, the model raised quality for non-experts. Used as a ghostwriter, it delivered no significant benefit — and was "in fact detrimental to expert users," via an anchoring effect that pulled their thinking toward the model's first suggestion.
Handing the model the pen is the failure mode. Any plan that has AI running creative strategy with no human in the loop is describing exactly the condition that made expert output worse.
Chen & Chan, "Large Language Model in Creative Work: The Role of Collaboration Modality and User Expertise," Management Science 70(12), 2024board
writer
A randomised field trial at a consulting firm found the same thing from the other direction: AI raised creativity especially for employees with strong metacognitive strategies — the habit of thinking about their own thinking. Tool access alone did nothing.
Sun, Li, Foo, Zhou & Lu, Journal of Applied Psychology, 16 June 2025 (RCT, 250 employees, supervisor plus two blind external raters)AI skills now carry a 28% salary premium — about $18,000 a year — across 1.3 billion job postings. In marketing and PR, AI-skill demand is growing 50% annually.
PwC's analysis of close to a billion job ads puts the premium higher still, at 56%, up from 25% the year before — and finds the skills employers ask for changing 66% faster in the most AI-exposed roles.
Lightcast, "Beyond the Buzz," July 2025 (1.3B job postings) · PwC, "2025 Global AI Jobs Barometer," June 2025
Photo: Nicolas Hoizey / UnsplashTHE LADDER
PROBLEM
If taste comes from years of doing the work, and AI is eating those years — where does the next generation of taste come from?
Nobody is born with taste. It gets built by making a lot of mediocre work and having someone senior tell you why it's mediocre.
Every study in the last chapter points to the same conclusion: judgment is the variable that decides whether AI lifts your work or quietly flattens it. But judgment is not a tool you can buy or a prompt you can write. It's an output — of years spent on unglamorous production work, under correction, close to people who already have it.
Which raises the question this chapter is about. That apprenticeship is the exact tier of work AI is absorbing first — and, as the labour data shows, the exact tier that is disappearing.
Stanford researchers went into the payroll records of the largest US payroll provider. The entry rung is being removed.
Workers aged 22–25 in AI-exposed occupations saw 16% relative employment declines, while employment for experienced workers in the same occupations held steady. Marketing and sales roles show the same shape.
The authors' mechanism, in their words: "task substitution at the apprentice margin, complementarity at the expert margin." Firms aren't firing seniors. They're not hiring juniors.
Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?", Stanford Digital Economy Lab, Nov 2025 (ADP payroll microdata)The people who teach taste are managers. Management is being deleted at speed.
Amazon told every S-team organisation to "increase the ratio of individual contributors to managers by at least 15%," stating plainly: "Having fewer managers will remove layers and flatten organizations."
Andy Jassy, Amazon CEO, official Amazon newsroom, 16 Sept 2024Manager headcount at public companies fell 6.1% between May 2022 and May 2025. Estée Lauder and Match Group each cut around 20% of their managers.
Live Data Technologies, reported via CNBC and The Wall Street Journal, 15 July 2025Put it together and you get the real risk in this report. Not that AI replaces creative people — but that it removes the junior work and the senior mentors that together manufacture taste, in the same decade that taste becomes the whole job.
Photo: max fuchs / Unsplash91% of marketing teams are increasing content output. 46% are producing three to five times more than in 2024. 75% got a budget increase of between one and ten percent.
10Fold with Sapio Research, Aug 2025 (n=400 senior marketing executives across US, France, Germany, UK; ±4.6% at 95% confidence)
Photo: Eric Prouzet / UnsplashBUILD THE
STANDARD IN
Three moves for leaders who'd rather not find out what a team with no taste produces.
Hire and promote for judgment, not throughput.
When execution is cheap, output volume stops being a signal of anything. The evidence says the variable that decides whether AI adds or subtracts value is the human's ability to evaluate what it produced — metacognition in the field trial, domain expertise in the wall-effect study, baseline performance in the entrepreneur study.
Interview for that. Ask candidates to critique work, not just produce it.
Across 1.3 billion job postings, roles asking for AI skills pay about $18,000 more a year. In marketing and PR, that demand is growing 50% annually — and PwC puts the premium higher still, at 56%.
Lightcast, "Beyond the Buzz," July 2025 (1.3B job postings) · PwC, "2025 Global AI Jobs Barometer," June 2025"Quickads helped us translate 150 years of craftsmanship into creative that actually performs — without losing the soul of the brand."
Protect the apprenticeship, or don't expect seniors.
A 16% relative decline in early-career employment and a 25% fall in promotions are not someone else's problem to solve in five years. They are your future creative directors, not being made right now.
Use AI to remove the drudgery from junior roles — not the roles. Keep juniors in the room where work gets judged, because that room is the only place taste has ever been transmitted.
Put the standard in the system, not just in people's heads.
If taste is now the constraint, treat it as infrastructure. Encode what good looks like — performance data, brand rules, what actually converted — so the standard survives when a person leaves and holds when volume triples.
It's demonstrably doable: predictive creative scoring has forecast 3-second view-through rate with 83% accuracy and contributed an 11% increase in profit ROI on scored assets.
VidMob with Kellanova and MMA Global, Aug 2025 (443 creative assets, Meta, US)"Quickads turned our ad workflow from a grind into a growth loop."
This is the design principle behind Quickads: 30M+ ads of performance intelligence encoded into the system, 500K+ ads generated for 30,000+ brands across 120 countries, and strategists who've run $100M+ in spend — so scale doesn't cost you the standard.
Quickads published company data, quickads.aiEVERYTHING IS ABOUT
TO LOOK THE SAME.
That's not a prediction — it's a measured finding, from Science Advances, from PNAS Nexus, from four million pieces of real creative work. AI lifts every individual and narrows the collective. In a market converging on one aesthetic, the only durable advantage is the judgment to leave it.
Almost every creative leader will tell you the work has to stay distinctly human. They're right. The catch is that staying human is now something you have to actively fund, protect and build into the system — because the default setting is everything converging on the same middle.