The Average Machine

Four findings from the recent research literature, each independently replicated, add up to one uncomfortable conclusion.

Regression to the mean is structural, not a bug to be prompted away. A 2025 paper applied directly to advertising, titled after "Galton's Law of Mediocrity," puts it plainly: because language models are trained to maximize the likelihood of the next word, they favor high-frequency patterns and suppress rare ones. Their outputs converge on safe, predictable phrasing, and the authors argue the pull gets stronger as training data grows, since a bigger corpus means a stronger center to regress toward.^[1]

Novelty and quality trade against each other. Work from New York University published in 2025 mapped what researchers call a frontier: the tricks available at the point of use, such as clever prompting or feeding the model fresh examples, tend to buy originality at the cost of quality. The frontier itself only moves with a stronger base model or targeted retraining, which means "just prompt it better" has a hard ceiling.^[2]

The training that makes models polite makes them predictable. A 2026 study evaluating creativity with human judges found that the largest, most advanced model families sit at the most conservative end of the scale, scoring high on appropriateness and low on novelty, while smaller, simpler families often score as more creative.^[3] The researchers' explanation, offered as a hypothesis rather than a proven mechanism, is that instruction tuning and safety training shift models toward safer, more predictable output. Earlier work on the training method behind aligned chatbots points the same way.^[5] For a creative department, this is the finding to sit with: the best-behaved model in the room may be the worst ideator.

AI assistance homogenizes work across people. A 2024 study by Doshi and Hauser found that writers using AI help felt more creative and produced individually better-rated stories, while the pool of stories across all writers became measurably more alike.^[4] Each person gets a lift. The industry as a whole converges. If every agency's ideation runs through the same handful of models, every agency's ideas drift toward the same center, and the drift is invisible from inside any one building.

None of this says the technology is useless. It says the technology has a shape, and the shape is a funnel toward the middle.

Reverse-engineering the last four years of Cannes Grand Prix winners surfaces a short list of recurring mental moves. What is striking is how repeatable they are, and how consistently each one involves rejecting the answer the brief implied.

Inverting the brief. Channel 4's "Considering What?" for the 2024 Paralympics (4Creative, one of the two Film Grands Prix at Cannes 2025^[8]) is the cleanest recent case. Channel 4's own research found 59% of viewers watched the Paralympics to see athletes "overcoming their disabilities" and only 37% for the sport.^[7] The brief implied another triumphant hero film. The team instead made the patronizing viewer the problem, personifying gravity, friction and time as hecklers who could not care less about disability. The channel's marketing lead called abandoning the equity-rich "Superhumans" platform scary. That fear is usually the tell that the inversion is right.

Breaking a category rule on purpose. L'Oréal Paris put out a 17-minute documentary, "The Final Copy of Ilon Specht" (McCann Paris, the other of 2025's two Film Grands Prix^[14]), about the dying copywriter who wrote "Because I'm Worth It" at 23. It plays as cinema, with no product pitch anywhere in it. McDonald's ran "Raise Your Arches" (Leo Burnett London, 2023), reportedly the first McDonald's film with no restaurant and no food, only a raised-eyebrow invitation. System1, the company that measures emotional response to advertising, scored it highest of the Cannes Film winners it tested that year.^[9] In both cases the violated convention was the idea.

Reframing the problem inside a specific culture. The New Zealand Herpes Foundation's "Best Place in the World to Have Herpes" (Motion Sickness and Finch, Grand Prix for Good and the Lions Health Grand Prix for Good, 2025^[12]) took a destigmatization brief that every health body on earth answers with solemn empathy and answered it with a parody tourism campaign fronted by a legendary All Blacks coach, aimed squarely at New Zealand's national pride. The foundation's own site carries the fact that made the humor land: up to 80% of people will carry the virus in their lifetime.^[13] Cadbury's "Shah Rukh Khan My Ad" (Ogilvy India, Creative Effectiveness Grand Prix 2023,^[10] later ranked first in WARC's Effective 100^[11]) started from a culturally precise observation: small local stores across India were collapsing after the pandemic, and Cadbury's future depended on them. The idea was to gift the brand's biggest asset, its celebrity, to more than 2,500 small businesses, from corner grocers to electronics shops.

Shipping infrastructure instead of messages. FCB Chicago's "Caption with Intention" (three Grands Prix plus a Titanium Lion, 2025^[15]) redesigned closed captioning itself: synchronized word-by-word timing, color-coded speakers, typography that carries tone. The Academy of Motion Picture Arts and Sciences is a partner on the project, and the agency reports the system has been adopted for Oscar-submission captioning. "Backup Ukraine" (Virtue, the Danish UNESCO Commission and Polycam, 2022 Digital Craft Grand Prix^[18]) shipped a 3D-scanning tool so citizens could digitally preserve heritage under bombardment. The campaign is the artifact, and the artifact outlives the media plan.

Intervening in a platform's mechanics rather than buying its media. Dove's "Real Beauty Redefined for the AI Era" (Mindshare, Media Grand Prix 2025^[16]) worked with Pinterest to retrain its recommendation algorithm toward a wider distribution of beauty, with women generating and pinning the training content themselves. The media buy was, in effect, a model-tuning intervention.

Joining a conversation already in motion. Vaseline found creators publishing thousands of TikTok "hacks" using its product, some brilliant, some dangerous. Instead of commissioning influencer content, "Vaseline Verified" (Ogilvy Singapore, two Grands Prix plus a Titanium Lion, 2025^[17]) lab-tested the existing hacks, awarded the ones that worked and put the failures on billboards.

The pattern across all of them: the breakthrough lives in what the brief was understood to be asking, before a single execution existed. Each winning team found the moment where the obvious answer was the trap. And the obvious answer is precisely, mathematically, what a language model is trained to produce.

Here is the state of the record, checked against Cannes Lions, D&AD, the Effies and WARC.

Where AI appears in awarded work, it appears as production technology inside a human-authored idea. Rephrase.ai's face and voice synthesis made Cadbury's thousands of personalized small-business ads possible, and the insight, the gifting idea and the Diwali timing were authored by people at Ogilvy and Wavemaker.^[10] A juror on the 2025 Creative Effectiveness Lions described a paradox: AI was everywhere on the festival stages and nearly absent from the effectiveness evidence in the winning work.^[24]

The one unambiguous AI headline from Cannes 2025 ran the other way. DM9's "Efficient Way to Pay" for Whirlpool's Consul brand lost its Creative Data Grand Prix after the festival found the case film used AI-generated and manipulated footage, including altered CNN Brasil news broadcasts edited to look like coverage of the campaign. The agency's co-president resigned.^[19] Cannes responded with a full integrity regime for 2026: mandatory disclosure of AI and synthetic media in every entry, a dual-layer verification system that scans submissions, senior sign-off from both agency and brand, sanctions up to three-year bans, and an independent Integrity Council of legal, ethics and industry experts to rule on escalated cases.^[20]

The festival simultaneously opened the door and drew the line, adding AI Craft subcategories across Design, Digital Craft, Film Craft, Industry Craft and Creative Data, framed in the festival's own words as recognizing "the craft and artistry of work that couldn't exist without AI but doesn't just use it as a tool."^[21]

Then June 2026 arrived, and the results are worth reading closely. The first AI Craft Grand Prix went to Google's "Project Genie," a consumer interface for its own world-generating AI model: a technology company awarded for crafting with its own technology.^[22] The Film Grand Prix, the prize that most defines the industry's idea of a great ad, went to Mother London's Super Bowl work for Claude, Anthropic's AI assistant.^[23] An ad about AI, conceived and made by humans. Jurors spent the week saying versions of the same sentence: in the AI era, craft and human judgment matter more, and the new disclosure rules were in force for the first time.

Read the scoreboard honestly and it says: after three years of the loudest technology adoption in the industry's history, and one full festival cycle with AI categories open, the machine's confirmed contributions to the work that defines the craft are execution, scale and one integrity scandal. The ideas remain human, so far without exception.

There is a live question underneath all of this: can the accumulated wisdom of the industry, its effectiveness data and planning frameworks and jury standards, be encoded into AI at all?

The answer from the research is yes, with a sharp condition. Frameworks translate well when used to judge work and badly when used to make it.

Consider what the diagnostic half of the industry's knowledge looks like. Binet and Field's effectiveness data on the balance of brand building and activation. Orlando Wood's taxonomy of the features that make advertising broadly appealing: characters in relation, a sense of place, story, humor, a repeatable creative device. System1's measurement of emotional response. The behavioral science catalog of Sutherland and Shotton. Cannes jury criteria themselves. All of these are explicit, codified and testable. A model can apply them to a script or a cut and produce a genuinely useful reading: this work has no characters, no place, no story; this idea is the category default; this "insight" is a purpose-anthem cliché wearing a lanyard.

Now hand the same frameworks to the model as instructions for generating and watch what happens. Ask for "a Bill Bernbach idea" and you get Bernbach cosplay. Ask for an idea that follows Binet and Field and you get a competent anthem. The frameworks describe what winning work has in common after someone had the idea. They do not contain the idea. Feeding them in as a recipe produces work that resembles an ad without ever being one.

The multi-agent research adds a warning for anyone building internal tools: naive setups where several AI agents discuss and refine ideas together produce more homogeneous output, because the agents talk each other toward consensus. What measurably helps is structural diversity. A peer-reviewed 2025 study of multi-agent ideation, run on scientific research ideas rather than advertising, found that a small panel of around three specialized critics with two or three rounds of revision beat both a lone model and larger, longer setups, and that diversity on the critic side mattered more than diversity among the generators.^[6] Even then, these systems improve the floor. Nothing in the published record shows one raising the ceiling.

So the line is drawn. On one side, the machine as critic: encodable, useful, honest. On the other, the machine as originator: structurally pulled toward the average, dressed in the language of your favorite dead genius.

For a working creative team, the research supports a handful of concrete practices.

Make the machine name the lazy answer, then ban it. The most valuable thing a model can do with your brief is complete it in the most probable way, on purpose, labeled as such. Ask it for the ten most predictable answers to the brief and treat that list as the forbidden zone. You now have a map of the trap, drawn by the trap.

Use it as a critic panel, and make the critics specific. A vague "give me feedback" invites politeness. Ask instead for separate, independent reads: where does this idea follow category convention, which parts are broadly appealing craft and which are abstract claims, what would an effectiveness jury question, where is the purpose-anthem cliché hiding. Run the critiques separately rather than as one conversation, since models reviewing each other's reviews converge on agreement.^[6]

Never let it supply your reference cases. A model asked for "campaigns like this" will confidently invent or misremember them, and invented precedents are exactly how average thinking sneaks back in wearing a Lion. Pull real cases from real archives. The same discipline applies tenfold to results: after 2025, fabricated case material is a career-ending category of mistake, and the festival now scans for it.^[20]

Keep the divergent phase wide and the convergent phase human. The defensible split, and the one worth testing inside your own team, is that the machine explores and the human decides. Thirty reframes of the brief, a hundred throwaway starters, fifty critiques: cheap, fast, useful. Which insight is alive, which idea is brave, which craft is right: those calls carry the taste the machine demonstrably lacks, and they are the job.

Apply the brief-inversion test to any tool you are sold. One question separates a creative tool from a production tool. Given the brief, can it tell you the brief is wrong and argue for a better one? Almost nothing on the market can, because almost everything on the market is trained to complete the assignment. Completing the assignment is the average machine's whole nature. Refusing it well is still ours.