Competent and forgettable is a worse commercial outcome than rough and distinctive. It is also much harder to notice from the inside.
There is a real experimental finding underneath this, and it is worth stating precisely because it gets mangled constantly. Writers given AI-generated story ideas produced work judged more creative and better written, with the biggest lift going to the writers who had scored lowest on a creativity measure beforehand. Those same stories were measurably more similar to one another than the ones written without AI. So it lifts the floor and narrows the range at the same time. For anyone producing creative work commercially, that means the advantage moves from execution quality, which is now cheap, to whatever you bring that the model could not have: what happened in the room, what the customer actually said, and what you decided to leave out.
Third round of concept review on a B2B campaign. Four routes on the wall, from three different people who hadn’t seen each other’s work. And they had the same skeleton: an uncomfortable industry truth, a reframe, a three-word line underneath.
Individually every one of them was fine. Better than fine. Tighter than what that group would have produced two years earlier, and you could see the craft in them.
Put side by side they were one idea in four outfits. Nobody had copied anybody. They’d all just started in the same place, with the same tool, from a brief that gave the model roughly the same handful of facts.
That’s the thing this article is about, and the reason it’s hard to catch is that it doesn’t show up in any individual review. Your draft looks good. It is good. The problem only exists at the level of the category, and nobody’s job is to look at the category.
Bring the thing the model can’t have: what actually happened in the room.
For anyone selling something, this is a commercial problem before it’s a creative one. Competent and forgettable is a worse outcome than rough and distinctive, because forgettable work costs you exactly as much to produce and buys you nothing on the way out. It also passes review more easily, which is the trap.
I’m not going to argue that anyone should use less AI. I use it heavily and my work is better for it. The argument is narrower: the value has moved. Execution quality used to be a differentiator and now it’s table stakes, so the differentiator has to come from somewhere else.
There’s one experiment that gets cited constantly for this and mangled about half the time, so it’s worth stating carefully.
Anil Doshi at UCL and Oliver Hauser at Exeter ran an experiment with 300 participants writing very short stories. One group got no AI help. One could take a single AI-generated starting idea. One could pick from up to five. Six hundred other people, who didn’t know which group anything came from, then rated the results.[1]
The stories written with AI ideas were rated more creative, better written and more enjoyable. Not less. If you’ve seen this cited as evidence that AI damages creativity, that’s the opposite of what it found.
Two details make it more interesting than the headline. The lift landed almost entirely on the writers who had scored lowest on a creativity measure taken beforehand, which effectively closed the gap between them and the strongest writers. And the strongest writers got no measured benefit at all.
Then the part that matters here. When the researchers compared the stories to each other rather than judging them individually, the ones written with a single AI idea were about 10.7% more similar to one another than the no-AI group.[1]
So both things are true simultaneously. Better individually, more alike collectively. The authors describe it as resembling a social dilemma, and that framing is the useful part: each individual writer is right to use it, because their own work improves, and the aggregate outcome is a narrower range of work than existed before.
Stories written with a single generative AI idea were about 10.7% more similar to each other than stories written without one, while the writers who already scored highest on the study’s creativity measure showed no judged improvement. Source: Doshi and Hauser, Science Advances vol. 10 no. 28 (2024), as described in the University of Exeter release. [1]
Two things to hold against that number before anyone builds a strategy on it. It’s 300 people writing eight-sentence stories on a research platform, which is a long way from a marketing team working on a campaign with a client, a category and a history. And the full paper sits behind a paywall from where I am. What I’ve quoted comes from the published abstract and the authors’ own university release. If a figure isn’t in this piece, it’s because I couldn’t read it myself.
Take it directionally. The direction is enough, because it matches what’s visible on any category’s LinkedIn feed right now.
When people worry about this they picture the writing stage, which is the one place it’s easiest to spot and probably does the least damage. The earlier three matter more, because everything downstream inherits them.
| Stage | How it gets in | The tell | What to change |
|---|---|---|---|
| 1. The brief | You ask it to tighten your brief before anyone has argued about it. The tightened version is the industry-standard version. | Your brief could be handed to a competitor with three nouns swapped and still make sense. | Write the messy brief first, in your own words, including the part you’re unsure about. Tidy it afterwards, if at all. |
| 2. The strategic frame | You ask for angles on a category. It returns the well-documented ones, because those are the ones that exist in text. | The angle feels immediately right and slightly familiar. That combination is the warning. | Ask it to argue against your frame instead of generating frames. Rejection is a better use of it than ideation. |
| 3. The structure | Everyone’s outline converges on the same shape: tension, reframe, proof, line. | Four routes from three people share a skeleton, which is exactly what happened in the story above. | Decide the structure before you open anything. Even a bad structure you chose beats a good one you inherited. |
| 4. The sentences | The polish stage smooths out the odd phrasings, and the odd phrasings were the voice. | It reads well and you can’t remember writing any specific line in it. | Keep one deliberately awkward thing per piece. It’s usually the bit people quote back at you. |
The four entry points, in the order they happen. Ours, from working on campaigns rather than from published research. The earlier ones are cheaper to fix and do more damage if you miss them.
Stage two is where I’d concentrate. A model’s suggestions come from what has been written down, so on any established category it returns the consensus positions with unusual fluency. That’s genuinely valuable for finding out what the consensus is. It’s a bad way to find out what nobody’s saying.
Stage four gets more attention than it deserves, though the tell is a good one. If you read something back and can’t remember making any particular decision in it, that’s usually because you didn’t.
Worth saying who this hits differently, because “marketers” covers people with very different exposure. A content marketer producing volume feels this at stages three and four, where the pull toward a template is strongest. A brand strategist feels it at stage two, and it’s more dangerous there because it’s invisible and it sets everything after it. A performance marketer writing ad variants arguably shouldn’t worry much at all, since convergence on a proven structure is the job. Different roles, different stage to defend.
Here’s the reframe that changed how I work. The model isn’t short of language. It’s short of your specifics.
It has never sat in your customer’s warehouse. It doesn’t know that your best account came from a complaint, or that the sales team has stopped using the phrase in your positioning because it makes prospects wince. Everything in that category is unavailable to it unless you put it there, and it is exactly the material that makes work impossible to confuse with anyone else’s.
Which means most of the sameness problem is an input problem wearing a creativity costume. If two people give a model the same three generic facts, of course they get similar work back.
| What to bring | What it looked like on one campaign |
|---|---|
| A sentence somebody actually said | Ops director, in a call, unprompted: “We don’t need better reporting, we need fewer people asking me for reports.” |
| A specific moment | Their team does a manual reconciliation every Thursday afternoon that everybody hates and nobody has ever put on a slide. |
| A number only you have | Eleven of the last fourteen closed deals mentioned the same competitor by name at the same stage. |
| Something that went wrong | The campaign last spring that tested well and did nothing, and the one line in it we still think was right. |
| A thing you refuse to say | No “transform.” No “seamless.” No claims about time saved that finance can’t stand behind. |
| What you already believe | Your actual position, written before you ask for input. So you can tell whether the output changed your mind or just replaced it. |
A filled-in raw-material brief. The last row is the one people skip, and it’s the only one that lets you notice when your thinking has been overwritten rather than sharpened.
That last row is worth dwelling on. If you write your own position down first, you can compare. If you don’t, there’s no version of your thinking left to compare against, and being talked out of something you never articulated feels identical to being helped.
None of this needs to be long. Six lines. The difference between six lines of real material and a generic brief is most of the distance between work that could only be yours and work that could be anyone’s.
If you change one thing after reading this, change when you open it rather than how you prompt it.
Early use means you go to it for the idea, the angle, the direction. Late use means you’ve made those calls yourself, badly and in the wrong order, and then bring it in to pressure-test, sharpen, find the hole and cover the ground you’re bored of covering.
Use it on your material, not instead of it.
Early use isn’t wrong everywhere, and I want to be careful not to turn a useful heuristic into a rule about the whole world. On a task where convergence is the goal, early is fine. Fifteen subject-line variants on a proven structure, a first pass at a competitor landscape, the third version of a product description: bring it in whenever you like. Nobody is buying originality there.
Where it costs you is anything meant to be identifiably yours. Positioning, a point of view, a founder’s voice, the campaign idea itself.
| Stage of the work | Bringing it in early | Bringing it in late |
|---|---|---|
| Finding the angle | You get the documented angles, fluently, and they feel like discovery | You get your angle stress-tested, which is a smaller job and a better one |
| Structuring it | The default shape, which is also everyone else’s default shape | Your shape, checked for whether it actually holds together |
| Writing it | Fast, clean, generic. Very hard to add voice back in afterwards | Slower first draft in your voice, then genuinely useful line editing |
| Checking it | Nothing to check against, since it wrote the thing it’s assessing | An independent read of something it didn’t produce. This is where it’s strongest |
The same four stages, early versus late. The bottom row is the underrated one: a model reviewing work it did not write is a far more reliable critic than one marking its own homework.
The cost is real and worth naming: the late approach is slower at the start. You’ll produce a worse first draft. If you’re measured on volume, that’s a genuine tension and pretending otherwise is useless. The practical compromise most people land on is triage. Decide which pieces are meant to be distinctive, protect those, and let the rest converge.
And since this puts AI directly into work you’ll be judged on, it’s worth writing down who owns what before you start rather than after a line goes out that nobody quite remembers approving.
| What AI does | What you still own | How it gets checked |
|---|---|---|
| Argues against your angle and lists what a sceptical buyer would object to | Whether the objection is real or just conventional, and whether you’re changing course | You write your position down before you ask. If you can’t say what changed your mind, nothing did. |
| Line-edits a draft you wrote, for rhythm, length and repetition | Every phrase that carries voice. Restore anything it smoothed that you’d have said out loud | Read the edited version aloud. The moment you hear a sentence you wouldn’t say, put yours back. |
| Checks a finished piece against the category for how familiar it sounds | The decision about which similarities are fine and which ones are you disappearing | One person outside the team reads it cold and says who they think wrote it. |
A worked example of the split on distinctive work. The third row is the cheapest quality check available and almost nobody runs it.
Everything above is now something you know. Knowing it holds for about a fortnight, which is roughly how long any new resolution about process survives contact with a calendar.
Actually, let me say that more precisely, because the vague version isn’t useful. What decays isn’t the knowledge. It’s the practice, and it decays under specific, predictable conditions: a deadline, somebody asking where the draft is, and a blank document at four in the afternoon. Nobody abandons a principle. They just open the tool first this once.
Which is why the fix has to be structural rather than attitudinal. Three things, and none of them require anybody to be disciplined at four in the afternoon.
If a competitor could publish it unchanged, it isn’t yours yet.
The third one is the one that catches drift, and it has to be somebody else. You can’t run it on your own work, because you know what you meant, and knowing what you meant makes generic writing read as clear.
Teams put versions of this in place on their own all the time. What an outside session tends to add is the uncomfortable comparison itself: an hour spent putting your last quarter’s output next to two competitors’ with nobody in the room defending anything. That’s the part that’s hard to schedule for yourself, and it’s a chunk of what our workshops with marketing teams actually do.
So: a way to actually check, on work you’ve already published, rather than a feeling that things have gone a bit flat.
Take five pieces from the last quarter. Not your favourites, the ordinary ones. Score each against these six, one point per yes.
| # | Question | Score 1 if |
|---|---|---|
| 1 | Does it contain a fact only your organisation could know? | A number, a moment, or a sentence somebody actually said. Not a public statistic. |
| 2 | Is there a position in it a reasonable person could disagree with? | Somebody credible could argue the other side. If not, you’ve published consensus. |
| 3 | Could a competitor publish this with only the logo changed? | Score 1 if they couldn’t. This is the one that usually costs people the point. |
| 4 | Is there a sentence you’d recognise as yours out of context? | One line with your fingerprints on it. Awkward counts. Awkward often wins. |
| 5 | Does it say what you’re not claiming, or where this doesn’t apply? | A named limit. Models rarely volunteer these, so their presence is a good sign a human decided something. |
| 6 | Did the structure come from a decision rather than a default? | You can say why it’s in this order. “It’s how it came out” scores zero. |
Six questions, one point each, scored across five recent pieces. Ours, built from reviewing marketing output rather than from published research. Question three is the one worth arguing about in the room.
Reading the scores, roughly:
The most useful output of this isn’t the score. It’s the pattern in which questions you lose. Teams that consistently lose one and four have an input problem, which is fixable in a fortnight with the raw-material brief. Teams that lose two and five have a courage problem, which is a management question and not a creative one, because somewhere in the approval chain a person is removing the sharp parts.
Where you go from here, if you want the ideas side rather than the output side, our piece on training AI on your brand voice covers keeping the sound of the work yours, and there’s a related argument about holding voice consistent as output scales.
The smallest useful thing to do this week: take the last thing you published, and try to find one sentence in it that only your company could have written. If you find it quickly, you’re fine. If you have to go looking, you’ve learned something more useful than any audit would have told you.
Not individually, which is the part that surprises people. In the Doshi and Hauser experiment, stories written with AI-generated starting ideas were judged more creative, better written and more enjoyable than stories written without them. The effect was concentrated among writers who had scored lower on a creativity measure taken beforehand; the strongest writers showed no measured benefit. What changed was the relationship between the pieces rather than the quality of any one of them. Stories produced with a single AI idea were around 10.7% more similar to each other than the no-AI group. So the honest answer is that your work probably gets better and the pool of work gets narrower, and only the first of those is visible from inside your own drafts.
Because the two effects are measured on different things. Individual quality is judged piece by piece: is this well written, is it engaging, does it do something unexpected. Collective diversity is measured by comparing pieces to each other. A tool that reliably nudges everyone toward good, well-documented moves will raise every individual score while pulling the whole set closer together. The researchers describe this as resembling a social dilemma, which is a precise way of putting it: each person is individually correct to use it, because their own work genuinely improves, and the aggregate result is a narrower range than existed before. Nobody in that situation is making a mistake, which is exactly why it’s hard to fix by telling people to be more careful.
Later than most people do, and how much later depends on what the piece is for. On work meant to be identifiably yours (positioning, a point of view, a campaign idea, a founder’s voice), make the angle and structure decisions yourself first, then bring AI in to argue against them, find the hole, and line-edit. On work where convergence is genuinely the goal (subject-line variants on a proven structure, a competitor landscape, the third product description this week), bring it in whenever you like, because nobody is buying originality there. The cost of the late approach is real: your first draft will be worse and it will take longer. The compromise that survives contact with a deadline is deciding which tier a piece is in at kickoff, before the pressure arrives.
Change the inputs before you change the prompts. Most sameness is an input problem: if two people hand a model the same three generic facts about a category, similar output is the correct result rather than a failure. Six lines of material only you have will do more than any amount of prompt engineering. A sentence a customer actually said. A specific recurring moment in their week. A number from your own pipeline. Something that went wrong. A list of phrases you refuse to use. And your own position, written down before you ask for input, so you can tell afterwards whether the output sharpened your thinking or quietly replaced it. That last one is the one people skip, and it’s the only one that lets you notice you’ve been talked out of something.
Be present. Everything that makes work impossible to confuse with a competitor’s comes from access rather than ability: the thing the ops director said unprompted on a call, the reconciliation everyone hates that has never appeared on a slide, the campaign that tested well and did nothing. None of that exists in text anywhere, so no model has it unless you put it there. The other one is deciding what to leave out. A strong piece of work is defined at least as much by the claim you declined to make as by the ones you did, and refusal is not something models are good at. Both are judgment applied to specifics, which is different from generating language, and it’s the part that hasn’t got cheaper.
The central finding here comes from Anil Doshi and Oliver Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” published in Science Advances in July 2024. It is reported here with some care, because it is frequently cited as showing that AI reduces creativity and it shows close to the opposite: AI-assisted stories were rated more creative, with the gain concentrated among less creative writers, while being more similar to one another. The full text sits behind a paywall from where this was written. What is quoted comes from the published abstract, held in UCL’s open-access repository, and from the University of Exeter’s own release describing the study, both read on 27 August 2026. A widely-indexed summary page carrying different figures for the same experiment was found and deliberately not used, because that site states its annotations are partly machine-generated and unverified, and its numbers disagree with the authors’ own institution. One sentence in the university release contains an evident duplication error, and it is not repeated here. Everything about where sameness enters marketing work, the raw-material brief and the drift audit is Future Factors’ own, drawn from producing and reviewing creative work rather than from research, and is labelled that way in the text.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.