AI answer convergence: whose strategy is this?

One question first

There are three ways to use a model on a hard question.

Almost everybody uses the first one, because it is the fastest and because nobody chose it deliberately. Pick the one that matches how you worked last week and read what it is costing you. The prices below are real and none of the three is free.

What it buys you

Speed, and a real lift in quality if writing is not your strength. In the Science Advances study, writers with the most access to AI ideas scored 8.1% higher for novelty and 9% higher for usefulness, and the least confident writers gained most.

What it costs

You inherit the model's defaults, and so does everybody else who opened with the same question. The same study measured a 10.7% rise in similarity between pieces of work that started from a single AI idea.

On a Tuesday

Your first draft exists in four minutes and is already the average of the internet. Everything after that is you editing your way back towards something of your own.

Fine when

The work is meant to be conventional. A policy summary, a job advert, the third version of a standard proposal. Convergence only hurts where difference was the point.

Most companies are on route one for everything, including the four or five decisions a year that actually separate them from the company down the road. The rest of this page is about how to tell which decisions those are and what to do differently on them.

How it turns up

Three plans, three companies, one sentence.

These are from three different companies in three different markets, in the same quarter. Read the verbs rather than the nouns.

A logistics scale-up, 140 people

We will differentiate on service quality by using AI to augment our people rather than replace them, building durable advantage over the long term.

A B2B software company, 90 people

Our edge comes from differentiated customer experience, with AI augmenting the team so we invest in long-term defensibility rather than short-term margin.

A healthcare services group, 320 people

We differentiate through quality of care, using AI to augment clinical staff and protect long-term outcomes over quarterly savings.

Differentiate. Augment. Long term. Three companies that compete with nobody in common, arriving at the same three words in the same order. Composites from client documents, lightly reworded so nobody is identifiable.

None of those sentences is wrong. Each one would survive a board meeting. And not one of them tells you what this particular company will do on Monday that the other two will not, which is the only thing a strategy is for.

AI answer convergence is everyone getting the same good answer, which makes it a worse answer. Across thousands of simulated strategy questions, leading models selected almost the same fashionable option regardless of the company in the prompt.1

The short version

Researchers at Esade, the University of Sydney and NYU Stern ran leading models through thousands of strategy dilemmas. The models kept choosing the trendy side, whatever the context. They named it trendslop.1
Better prompting and more context made less difference than you would hope.1
In a controlled writing experiment with 300 writers, AI help made individual work measurably better and made the whole set 10.7% more similar to itself.2
A separate study found the sameness comes from the model handing similar ideas to different people, rather than from any one person getting stuck.3
People in that study also produced more ideas, in more detail, and felt less responsible for them.3 That is the part with consequences.

You are here because something like this happened

Two plans from different teams came back sounding like the same document
Your strategy deck reads well and says nothing only you could say
A competitor announced roughly your idea, roughly when you had it
You suspect the tool is agreeing with the internet rather than with you

What you take away

Three ways to use a model, and the honest price of each
A two-prompt test that shows whether your context was ever in the answer
The research, with sample sizes and what is second-hand
The one thing to write down that a model cannot get anywhere else

Defining it

A model gives you the middle of everything ever written about your question.

That is what the thing does, and most of the time it is exactly what you want. Ask it how to structure a payroll change and the middle of everything ever written about payroll changes is a very good answer.

Ask it what your company should do next and you get the same treatment, which is where it goes wrong. There is no middle of everything written about your company, so it gives you the middle of everything written about companies that sound a bit like yours. Then it writes that up so well that it reads like a conclusion.

On strategy, LLMs might be more akin to a freshly minted MBA or junior consultant, parroting what's popular rather than what's right for a particular situation. Natalia Levina, NYU Stern, on the trendslop research1

The other half of the problem sits outside the machine entirely. When thirty-six people were given the same brainstorming task, the ones using ChatGPT produced ideas that were less different from each other than the ideas of the group without it. The researchers looked for the cause inside each person and did not find it there. Each individual's own set of ideas was about as varied as before. The sameness appeared between people, because the tool was quietly offering all of them the same starting points.3

So you cannot fix this by telling your team to be more creative. Every one of them is being creative in the normal way, on top of a starting point that thirty other people were also given.

01Campfire60%

02Wild West25%

03Blueprint10%Your own method

04Engine4%

05Ecosystem1%

Share of companies per level. At Level 2 everybody prompts their own way, so everybody lands on the same defaults. Level 3, Blueprint, is where a company writes down how it decides, which is the first thing a model cannot copy from the internet. The shares per level come from emaho's own work with clients.

This is where the culture level of a company starts to matter more than the tools it buys, and it is the reason two companies with identical subscriptions end up in very different places.

What the studies found

Everybody got better. Everybody got closer together.

The cleanest measurement of this comes from a controlled experiment published in Science Advances. Three hundred people were asked to write an eight-sentence story for young adults. One group worked alone, one could see a single three-sentence idea from ChatGPT, and one could choose from five. Six hundred other people then judged the results without knowing which was which.2

What it did for each writer

Compared with writing alone, judged by 600 readers

More novel+8.1%

More useful+9.0%

Better written, least confident writers+26.6%

Less boring, least confident writers+15.2%

What it did to all of them together

Similarity between the stories, measured across the set

More alike than the group with no AI+10.7%

One idea from one model, and three hundred people moved closer together. Nobody in the experiment could see this happening, because it only exists between people.

Bars are scaled within each side, so the two sides are not directly comparable. The left side shows the change in how work was judged, the right side the change in how alike the work became.2

Read those two sides together and you have the whole challenge. Everyone in the room got better at the thing. The room got worse at producing different things. If your job is writing one story, take the deal. If your job is running a company that has to be distinguishable from three others, look at the right-hand side again.

300

writers in the experiment, judged by 600 readers who did not know which stories had AI help

2

36

people in the brainstorming study, where the sameness showed up between users rather than inside any one of them

3

7

classic strategic tensions put to leading models across thousands of simulations, with the trendy side winning nearly every time

1

The strategy study is the one that should worry a leadership team most, and it is also the one to read with the most care. Reports of it name seven familiar tensions and say the models leaned consistently towards differentiation over commoditisation, augmentation over automation and the long term over the short term. Those specifics come from write-ups rather than from the article itself, which sits behind a paywall, so treat them as reported. What the authors state in their own summary is blunt enough on its own: across thousands of simulations the models almost uniformly picked the same fashionable strategies, regardless of context.1

And the finding nobody quotes

In the brainstorming study, people using ChatGPT produced more ideas and more detailed ideas, and reported feeling less responsible for them.3 Hold that next to a strategy meeting. A plan that nobody quite feels is theirs is a plan that nobody defends in month four, when it becomes inconvenient and somebody asks why you chose it.

One caution across all three studies. Two of them are small controlled experiments on writing and idea generation rather than on running a business, and the third is a simulation of strategic questions rather than a record of decisions companies actually took. They point clearly in one direction and none of them proves what happens in your Monday meeting.

The mechanism

Same models, same questions, same answer.

A good answer arrives before you have finished thinking

The old order was think, then write. Now something well-argued appears while you are still working out what you believe, and arguing with a fluent paragraph is much harder than filling a blank page. People know this about themselves and do it anyway, because the paragraph is right there and it is good.

The things that make you different are the things you never wrote down

Your model does not know that your biggest customer takes eleven weeks to pay, that your best engineer will leave if you move to the enterprise segment, or that the last time you tried this the founder overruled it. None of that is in any prompt, because none of it is in any document. So the answer is built on everything except the part that decides the outcome.

You asked about your company. It answered about companies. Paul Musters

Nobody is accountable for a plan they did not quite write

This is the one with the long tail. People who generate with a model feel less ownership of what comes out.3 In a strategy that means the plan survives the meeting and dies quietly in the execution, because when the first hard trade-off arrives there is nobody in the room who believes in it enough to pay for it.

Where does your company sit?

Six questions. Three minutes. Nobody sees who answered. Your level, the cost of it, and what changes at the next one.

Do the Culture Level scan

One thing to try

Ask the same question twice, ten minutes apart.

Take a real question you put to a model in the last month. Something that mattered, where you used part of the answer. Then run this, which takes about ten minutes and needs nothing you do not already have.

1

Ask it the way you asked it

Same wording, same context, same company details. Save the answer somewhere you can put it next to another one. Do not tidy the prompt up, because the sloppy version is the one you actually use.

2

Ask it stripped

New chat. Same question, with everything specific to your company taken out. No industry, no size, no history, no constraints. Just the question a hundred other people would ask.

3

Put them side by side

If the two answers say roughly the same thing in a different order, your context never reached the answer. You have been reading the generic version with your company's name typed at the top.

Almost everybody who runs this is surprised by how little the answer moves. What it tells you is where the specific things about your company live, which is in people's heads and nowhere the model can reach. That is a fact about your company rather than about the tool.

Do it once as a leadership team, on a screen, with a question one of you actually asked. It settles an argument that otherwise runs for months, and it takes less time than the argument.

Level by level

The difference you have left is in your people, and it is unevenly spread.

If the models give everybody the same starting point, what is left as an advantage is what your particular people do with it. That is not evenly distributed and it does not follow seniority.

An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.

Some people argue with a model by instinct. They read a confident paragraph and their first move is to look for what it left out. Others take a good answer as a finished answer, which is reasonable and much faster and is exactly how a company ends up with three plans that sound alike. Both types are useful and they need completely different things from you.

We read two things off each person: how they think and work, and where they are with AI. That tells you who to put on the four decisions a year where being different is the whole point, and who to leave on the work where the average answer is genuinely the right one.

About emaho

emaho measures one Operating Profile per person: personality type and AI level in a single profile. On that we build a personal set of AI agents that fit how that person works, inside the tools they already use. Fifteen minutes to complete, first profile free, built for companies between 20 and 500 people.

See your own profile, free Find your AI Culture Level

Fifteen minutes per person. No credit card, no strings.

What to do

Write down the things a model cannot know.

There is one first move here and everything else follows from it. It takes an afternoon, it produces a page, and that page is the only thing standing between your strategy and everybody else's.

Open a document and write what is true about this company that is written nowhere. Not the mission. The awkward operational facts: which customer you cannot afford to annoy, which two people the whole thing depends on, what you tried in 2024 that failed and why, what your board will never approve, where your margin actually comes from. Twenty lines is plenty. Most leadership teams have never had this page and discover halfway through that they disagree about line six, which is worth the afternoon on its own.

What that page then lets you do

It turns route three into something you can actually run. You paste those twenty lines in, give the model your decision, and ask it to argue against you using your own constraints. Without the page that request produces generic objections. With it, you get the objection your CFO was going to make in three weeks.

It also gives you a real test for any AI answer that arrives on your desk. Read it and ask which line of the page it used. If the answer would be equally true for a company that shares none of your twenty lines, it is a summary of the internet and you can treat it as one.

And one habit to build around it

Whenever a decision matters, have two people ask the question separately in their own words, then compare before anybody circulates anything. Where they match, the model has given you the consensus and you can move on. Where they diverge is where the actual thinking is, and that is the only part worth a meeting.

How much did your answer move?

If you run the two-prompt test, send me what changed between the two answers. I can usually tell from that which part of your context never made it onto paper. You get my read on it, and that is all.

Message me on WhatsApp

If you do nothing

You and your competitor walk into the same meeting with the same answer.

The cost you can put a number on arrives within a year. If three companies in a market all move towards the same fashionable position, they end up competing on price in the same place, because that is what is left when the difference goes. Nobody chose that. Each of them made a sensible decision on their own with a helpful tool.

Slower, and visible in your calendar: plans that nobody feels they wrote get executed with less conviction.3 You will see it as a strategy that keeps needing to be re-explained, and as decisions that get reopened every quarter because no one in the room ever really argued for them.

And there is the one that only becomes visible in hindsight. The research suggests a spiral: writers who find that AI-assisted work is judged more creative have every reason to use it more, and the more everybody does that, the less varied the whole becomes.2 Each individual choice is correct. The place they collectively arrive at is somewhere nobody would have chosen.

The common ones

Built from several conversations, not a single case.

What people ask when everyone gets the same answer

What is AI answer convergence?

It is what happens when many people ask the same AI models similar questions and receive similar answers, so the work produced across a market becomes less varied. Each individual answer is usually better than what the person would have written alone. The set of answers is worth less, because the difference between them has gone.

What is trendslop?

Trendslop is the term Angelo Romasanta, Llewellyn D. W. Thomas and Natalia Levina gave, in Harvard Business Review in March 2026, to AI's tendency to recommend fashionable ideas over reasoned ones. Across thousands of simulated strategy questions, leading models almost uniformly selected the same trendy strategies regardless of the company described in the prompt. The authors call the strategic version of it strategy trendslop.

Does AI make people more creative or less?

Both, at different levels. In a Science Advances experiment with 300 writers and 600 judges, stories written with the most AI help scored 8.1% higher for novelty and 9% higher for usefulness, and the least confident writers gained most: up to 26.6% better written and 15.2% less boring. The same experiment measured a 10.7% rise in how similar the stories were to each other. Individual creativity up, collective variety down.

Why do AI tools make everyone's ideas look the same?

Because the sameness happens between people rather than inside any one of them. A 36-participant study at the 2024 Creativity & Cognition conference found that people using ChatGPT produced ideas that were less semantically distinct from each other than those of a comparison group, while each individual's own range of ideas stayed about as varied. The model was quietly offering different users the same starting points.

How do you find out whether your company can produce a strategy a model could not have written?

Look at whether the way you decide is written down anywhere. emaho places companies on five AI Culture Levels, and writing down your own method is Level 3, Blueprint, where about one in ten organisations sits. Below that everybody prompts their own way, so everybody lands on the same defaults and the company has nothing of its own to put in the prompt. The scan is six questions, three minutes and anonymous.

Can better prompting fix AI answer convergence?

Only partly, according to the research. Write-ups of the 2026 trendslop study report that attempts to correct the bias with improved prompting and richer context had limited effect, because the models are reflecting prevailing internet discourse rather than reasoning about your situation. The more reliable fix is putting facts about your company into the prompt that exist nowhere online.

How do you know whether an AI answer actually used your company's context?

Run the same question twice. Ask it the way you normally would, then ask it again in a fresh chat with everything specific to your company removed: no industry, no size, no history, no constraints. Put the two answers side by side. If they say roughly the same thing in a different order, your context never reached the answer.

Is it a problem if competitors get the same AI advice?

It is, where difference was the point. If several companies in one market move towards the same fashionable position, they end up competing on price in the same place, and nobody decided that. For conventional work such as policy summaries or standard proposals, convergence costs you nothing and the speed is worth having.

Do people feel less ownership of ideas they made with AI?

Yes, and it is one of the more consequential findings. In the 2024 Creativity & Cognition study, participants using ChatGPT generated a greater number of more detailed ideas and reported feeling less responsible for them. Applied to strategy, a plan nobody quite feels is theirs is a plan nobody defends when the first hard trade-off arrives.

Where can I read about the other AI challenges inside companies?

This is one of 25 AI challenges emaho documents, each with the research behind it, where it sits in the five AI Culture Levels, and a test you can run this week. Answer convergence sits next to advice that is sound but fits a company you are not and a model that agrees with whoever is asking.

Which strategies do AI models tend to recommend?

Secondary reports of the 2026 trendslop research describe a consistent lean towards differentiation over commoditisation, augmentation over automation, and long-term over short-term performance, across seven classic strategic tensions. Those specifics come from write-ups rather than the paywalled article, so treat them as reported. What the authors state directly is that the models chose the same trendy strategies regardless of context.

How do you keep a strategy distinctive when everyone uses the same AI?

Write down what a model cannot know about you. Twenty lines of awkward operational fact: the customer you cannot afford to annoy, the two people the business depends on, what failed in 2024 and why, what your board will never approve, where the margin actually comes from. That page is what turns a generic answer into an argument about your company, and most leadership teams have never written it.

What is the best way to use AI on an important decision?

Give it your decision and ask it to argue against you, with your real constraints in the prompt. That uses the model for the thing it is good at, which is generating the case you have not considered, rather than for the thing it is weak at, which is knowing your situation. It also requires you to have decided something first, which is where most of the value is.

Does using AI for brainstorming produce more ideas?

More ideas and more detailed ones, according to the 36-participant study at Creativity & Cognition 2024. The catch is that the extra quantity did not come with extra variety: each participant's set of ideas was about as diverse as without the tool, while the sets produced by different participants overlapped more. More output, narrower range across the group.

Close by

Coming from the same place

All three are about what a model does with a question when it does not know anything about you.

13

Good advice, wrong company

The same missing context, seen from the other side. Sound advice that fits a company you are not.

Read this one 05

AI agrees with everyone

Why it rarely pushes back, what that does to the person asking, and the wording that gets you a real answer.

Read this one 22

Winners take more

What happens to the few companies that do not converge, and how quickly the distance opens up.

Read this one

All 25 AI challenges, and what to do about each

Twenty-five things that break inside a company once people start using AI, each with the research behind it and the level where it starts to bite. This one starts to bite at Blueprint, and 7 of the others start there too.

Open the full set

Getting going

Who in your company argues with the answer?

If everybody starts from the same place, your only remaining edge is what your particular people do next. Fifteen minutes each, and you can see who that is.

Run your last strategy question through three toolsIf the three answers rhyme, so does everybody else's in your market.
Find out who takes a good answer apartThe profile shows who does this by instinct. That habit is the one part a competitor cannot buy.

Start with your own profile, free Ask me anything first Plan a 30-minute call

First profile free · no credit card · built for companies of 20 to 500 · you decide what your team gets to see

Not ready to put your team in anything yet? Start with the level of the company instead. The Culture Level scan is six questions, three minutes, and asks nothing of you.

Paul MustersFounder of emaho

Paul Musters

Fifteen years of leadership and team development in Dutch scale-ups. That practice now sits in software: one Operating Profile per person, with agents that actually fit. He writes these pages from what he runs into with clients, not from a research summary.

LinkedIn · paul@emaho.world · WhatsApp

Sources and numbers used on this page

Angelo Romasanta (Esade), Llewellyn D. W. Thomas (University of Sydney) and Natalia Levina (NYU Stern), Researchers Asked LLMs for Strategic Advice. They Got “Trendslop” in Return., Harvard Business Review, 16 March 2026. Source of: the term trendslop, and the finding that leading models consistently recommend strategies aligned with managerial trends rather than context-specific logic, with the authors writing that across thousands of simulations the models almost uniformly selected the same trendy strategies regardless of context, and that on strategy an LLM may be closer to a freshly minted MBA or junior consultant parroting what is popular. The quotation and that summary are taken from the authors' own research highlight published by NYU Stern on 16 March 2026. The Harvard Business Review article itself is paywalled beyond its opening. Secondary write-ups describe seven classic strategic tensions and a consistent lean towards differentiation over commoditisation, augmentation over automation and long-term over short-term performance, and report that improved prompting and richer context had limited effect. Those specifics are second-hand and are flagged as such on the page.
Anil R. Doshi (UCL School of Management) and Oliver Hauser (University of Exeter Business School), Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, July 2024, DOI 10.1126/sciadv.adn5290. Details and quotations taken from the University of Exeter press release of 15 July 2024. Source of: 300 participants writing an eight-sentence micro story for a young adult audience, split into a group with no AI, a group given one three-sentence ChatGPT idea and a group choosing from up to five; 600 recruited judges assessing novelty and usefulness; stories by writers with the most AI access scoring 8.1% higher for novelty and 9% higher for usefulness; writers with the lowest scores on a Divergent Association Task gaining 10.7% in novelty and 11.5% in usefulness, and being judged up to 26.6% better written, up to 22.6% more enjoyable and up to 15.2% less boring; more creative writers benefiting least; and a 10.7% increase in similarity between stories written with one generative AI idea compared with the group that used none, measured with OpenAI's embeddings API. Professor Hauser describes the pattern as an emerging social dilemma in which individual incentives to use AI further reduce collective novelty.
Barrett R. Anderson, Jash Hemant Shah and Max Kreminski, Homogenization Effects of Large Language Models on Human Creative Ideation, Proceedings of the 16th Conference on Creativity & Cognition, 2024, DOI 10.1145/3635636.3656204, also available as arXiv:2402.01536. Source of: a 36-participant comparative user study in which different users produced less semantically distinct ideas with ChatGPT than with an alternative creativity support tool; the finding that homogenisation appeared at group level rather than as increased fixation within individuals; and the finding that ChatGPT users generated a greater number of more detailed ideas while feeling less responsible for the ideas they generated.

Numbers are quoted as published. The three plans in the second section are composites drawn from client documents, reworded so that no company is identifiable. The share of companies per level comes from emaho's own work with clients rather than from the studies above.