Ads made with AI: more of them, but all alike

Producing an ad has never taken so little effort. An assistant suggests thirty headlines in a minute, the ad platform generates images from your page, and a team that shipped five variants a week can ship fifty. Speed is no longer the interesting question. The interesting question is what happens when every advertiser in a sector speeds up with the same tools, trained on the same texts, and they all end up saying the same thing in the same way.

An experiment on 11,024 ads

Harang Ju and Sinan Aral ran a field experiment of unusual size. Participants, randomly assigned, worked either in pairs of humans or with an AI agent, and had to produce ads. These ads were then run in real conditions on a social network, for about 5 million impressions.

The result has two sides. Teams with an AI produced more, with better quality text. But their content was more homogeneous, and teams of two humans made better images. This is a working paper whose journal publication is not confirmed to date: a result made solid by its method and sample, but still to be confirmed by other teams.

The same effect, elsewhere

This is not an accident specific to advertising. Anil Doshi and Oliver Hauser published in Science Advances an experiment on short stories. Stories written with ideas from AI are judged more creative, better written and more enjoyable, especially among the least creative writers at the start, but they resemble each other more than stories written without AI.

We covered that study, and what it says about creativity in general, in our article on AI and creativity. What matters here is that both results point the same way: each person improves, the whole becomes more uniform.

Why ads end up alike

The mechanism is no mystery. A language model suggests the most likely continuation given what it is asked. If two competitors describe the same product with the same words, it will suggest similar hooks: the same benefits, the same verbs, the same promises of speed and simplicity.

In a news feed or on a results page, an interchangeable ad has a precise cost: it can be well written and leave nothing behind. The buyer sees three identical promises and chooses on price, or does not choose at all.

What does not look alike: your data

The only material a competitor cannot feed the same tool is your own facts. AI shapes them very well; it cannot invent them, and it must not try.

• Your operating figures. A real delivery time, a response rate under one hour, a founding year, as long as you can prove them.
• Your customers’ words. Objections heard in meetings, phrases from reviews, questions asked to support. They say what decides a purchase better than a generic pitch.
• Your campaign results. The ads that brought qualified enquiries, not just clicks, and the ones that failed.
• Your consenting contacts. A list to write to before paying to reach the same people elsewhere.

In practice, the brief given to the AI changes nature. Instead of “write an ad for our software”, you give it three real objections, a measured delivery time and the exact wording of a customer review, then ask for variants. It still produces fast, but from material nobody else has.

Volume, for testing rather than broadcasting

AI’s advantage on volume is real, as long as it is used to test, not to flood. Fifty variants are worth nothing if they all say the same thing; ten variants testing ten different angles, each fed by a fact of your own, teach you something.

1. 01Choose the angles.

   Three to five distinct promises, each backed by a verifiable fact about your business.

2. 02Let the AI produce variants.

   Several wordings per angle, from the data provided, adding nothing that is not in the brief.

3. 03Review what commits you.

   Prices, lead times, guarantees, comparisons: anything that could mislead is checked before it runs.

4. 04Test with enough volume.

   One angle at a time, judged on qualified enquiries, not clicks.

How many clicks it takes to reach a conclusion is the subject of our article on ad testing, and the review rules for generated copy and visuals are detailed in our article on AI ad copy and visuals.

We apply the same rule to our own content: the engine that prepares the articles on this blog only produces drafts, and none is published without human review, a rule adopted after a client had to take down from its site three pieces of content made up by AI.

A market about to fill up

Competition for paid attention is already wide. According to Eurostat, 32.6% of EU businesses with 10 or more people paid for internet advertising in 2024, and 27.8% in France (isoc_cismt database). As producing ads becomes almost instant for each of them, the number of ads grows faster than the attention available.

In this landscape, the work of the people who design campaigns does not disappear, it changes focus: less time writing the fifteenth variant, more time deciding what deserves to be said, finding the facts that prove it, and reviewing what the machine makes of it. That is the split our digital marketing offer is built on: AI prepares the variants, a person approves them. And to judge an ad on sales rather than clicks, sending CRM sales back to Google Ads remains the starting point. AI makes output abundant; difference becomes scarce again, and it comes from what you know about your customers.

Frequently asked questions

Sources

1. Ju and Aral, Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance, arXiv 2503.18238 (new tab)arxiv.org
2. Doshi and Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, 12/07/2024 (new tab)doi.org
3. Eurostat, isoc_cismt database (social media and advertising), extracted on 14/10/2026 (new tab)ec.europa.eu