Can you tell when a text was written by AI? What the research says

If you approve campaigns, you have probably heard that text generated by artificial intelligence goes unnoticed. Research published in recent years makes it possible to test that claim, and the results matter directly to anyone in charge of a brand's message.

In this second article of the series, we bring together three academic studies that measured how well people and statistical analyses recognize machine writing. The results vary according to who is reading, the type of text and how familiar the reader is with these tools.

People who use AI often recognize machine text

In 2025, researchers Jenna Russell, Marzena Karpinska and Mohit Iyyer published a study with a straightforward title: people who frequently use ChatGPT for writing are accurate detectors of AI generated text. They gathered 300 nonfiction articles in English, some written by people and some generated by commercial models such as GPT-4o, Claude and o1.

For your marketing, the most useful detail lies in the clues these readers used. They pointed to an "AI vocabulary", with words that show up unusually often in generated texts, and they assessed formality, originality and clarity. In practice, they noticed writing that sounds correct and predictable at the same time.

This group of readers is inside your company and outside it. Copywriters, journalists, content creators and communication professionals use these tools every day and have trained their eye without realizing it. Many of them are part of your campaign's audience and comment on social media about what they see.

Readers unfamiliar with the tool score close to chance

An earlier study tested readers without that daily exposure. In March 2023, Maurice Jakesch, Jeff Hancock and Mor Naaman published a study in the journal PNAS with about 4,600 participants, who evaluated 7,600 short self presentation texts, such as profiles of job candidates, dating app users and Airbnb hosts.

Participants identified the origin of the texts correctly 50% to 52% of the time, a result close to a coin toss. According to Cornell University's announcement, they relied on flawed shortcuts: mentions of family, life experiences and the use of the first person were read as signs of human authorship.

This result is often cited as proof that nobody notices AI. The design of the study helps you measure how far that conclusion reaches in advertising: the study tested short self presentation texts, with 2023 models, and untrained readers. A commercial faces greater scrutiny, because it circulates at scale, reaches trained readers and becomes a topic on social media.

Mor Naaman explained in the study's announcement that these systems produce statistically likely text, that is, common language. In advertising, this trait weighs even more, because your brand needs precisely to stand out from the common to be remembered.

The machine's fingerprints show up at scale

The third study looks at writing collectively. Dmitry Kobak and colleagues analyzed more than 15 million abstracts of biomedical papers indexed in PubMed and measured the use of words that grew abnormally after language models arrived.

The authors estimated that at least 13.5% of the abstracts published in 2024 were processed with language models, with rates reaching 40% in some subsets. The words that give this use away include terms such as "delves", "showcasing", "underscores" and "crucial".

For your brand, the study indicates that each model has vocabulary preferences, and these preferences leave measurable traces when many texts come out of the same tool. An advertising script born from this process tends to carry the same fingerprints.

What the divergence between the studies teaches you

Read together, the studies show that perception depends on who is reading. Accuracy was high among people who use AI often and stayed close to chance among untrained readers. Large scale analyses add patterns that a single text hides.

In a campaign, the part of the audience that notices is enough to create the problem. These are the people who comment, share the screenshot and set the tone of the conversation about the film. The third article of the series shows what happens when that conversation takes over a launch.

There is also a question that comes before detection. Even when nobody identifies the origin, the likely text tends to sound similar to so many others. NielsenIQ, as we saw in the first article, found that ads made with AI triggered weaker memory activation, including the high quality ones.