Can the audience tell when a speech was written by AI? What the research says

In an experiment published in the journal PNAS in 2023, 4,600 participants read 7,600 self-presentations, those short texts in which someone describes themselves, and tried to say which had been written by a person and which by an AI. They were right between 50% and 52% of the time, practically the result of a coin toss. "The machine passed this test," summarized Hancock, one of the authors, according to TechXplore. If you have ever read an AI text on stage, this finding may sound like a relief. The research that follows shows why that relief is short-lived.

The question in the title has several layers. The first concerns the ability to detect the text. The others concern what happens to your image when the audience suspects, when it finds out and when it notices that little personal effort went into the message. Here we gather the studies that answer each of them, along with the limitations each one carries.

Detecting AI text: people do about as well as a coin toss

The study by Jakesch, Hancock and Naaman analyzed how participants made their decisions. They relied on misleading cues, such as the use of first-person pronouns and the presence of family themes, which they associated with a human author. AI, which also writes easily about these subjects, slipped through the filter effortlessly.

The study has a limitation you need to know about: the texts were written self-presentations, read silently, and the experiment did not involve speeches delivered in front of an audience. Even so, the result dismantles a common idea among leaders, that there is an infallible detector in the audience's head. What weighs on the speaker, as the following studies show, is suspicion, and it is much harder to control.

Suspicion of AI weighs even when AI was not used

A study by Hohenstein, Kizilcec, Jung, Hancock and other researchers, published in Scientific Reports in 2023, observed conversations with AI-suggested replies. There were 361 participants in the first experiment and 291 pairs in the second. According to the summary published by PsyPost, those who believed their partner was using AI came to see them as less cooperative and more dominant, even in cases where the partner was not using any tool at all.

The same study found that the actual use of suggested replies improved the evaluation of the partner. The authors therefore draw attention to the weight of perception: the harm appears when the other person believes a machine took part in the conversation. "AI tools have social side effects," said Jung, according to PsyPost. We did not consult the primary source directly, and the numbers cited here come from that summary.

Now bring this to the stage. A speech full of sentences that seem to come out of a language model triggers suspicion, and suspicion hits your image regardless of who wrote the text. The fourth article in the series shows which constructions tend to set off this alarm in the audience.

The penalty for disclosing AI use in a message

Researchers at the University of Michigan, Zhu and Molnar, studied what happens when the reader is told that a message came from an AI. In the preprint version, with 647 participants in the United States and a preregistered design, 45.7% of participants said they had used AI to write messages in the previous two weeks. The central result: those who know a message came from an AI come to see the sender as lazy and insincere. Without that notice, no one suspects anything, and the impression is the same as that of a human text.

Keep in mind an important caveat. A later version of the work, published by Fast Company in partnership with The Conversation in April 2026, mentions more than 1,300 participants, and the journal in which the study was published could not be confirmed.

A paper presented at the CSCW 2023 conference by Karinshak, Liu, Park and Hancock reached a similar finding in another context. Participants rated pro-vaccine messages written by GPT-3 as more effective than those from the CDC, the US health agency, and expressed rejection of messages labeled as AI-generated. The study's abstract does not include numbers.

Personal messages: the speaker's effort is also judged

Liu, Kang and Wei, in a study published in the Journal of Social and Personal Relationships in 2023, asked 208 adults to evaluate messages from a friend. According to Ohio State University, the friend who used AI was seen as someone who made less effort, and satisfaction with the friendship dropped. Help from another person with the writing had a similar effect. "Sincerity and authenticity still matter a lot in relationships," said Liu.

For those of us who write speeches, the finding about human help carries a lot of weight. It shows that the audience values the speaker's personal investment. That is why, in the way we work, a speech script starts from your stories, your opinions and your way of speaking, gathered in conversation with you. The final text needs to carry marks that only you could have left.

Sermons, services and the audience that finds out at the end

Churches and synagogues have become a laboratory for this topic. In 2023, according to the AP news agency, Rabbi Joshua Franklin, of the Jewish Center of the Hamptons in the United States, asked his congregation to guess the author of a sermon. When he revealed that the text had been generated by AI, he was applauded and responded: "Now you are applauding. I am scared to death." He also wondered whether the tool is capable of empathy. In the same report, pastor and preaching professor Hershael York summed up his assessment of AI-written sermons: "It lacks soul." A member of another church described the AI text he heard as generic and strange.

In June 2023, at St. Paul's church in Fürth, Germany, more than 300 people attended a service generated by AI. Its creator, Jonas Simmerlein, estimated that about 98% of the content came from the machine, according to Greek Reporter. Churchgoer Heiderose Schmidt, 54, said that "there was no heart and no soul." Pastor Marc Jansen, 31, said he was pleasantly surprised. The source is secondary, and the original AP report on the service could not be consulted.

In both episodes, the audience receives the text attentively and then judges the experience by the criterion of human presence. It is the same criterion that a convention audience applies to the company president's speech, and that your team tends to apply to yours.

What the research tells you before you step onto the stage

Taken together, the studies paint a coherent picture. People are rarely right when they try to identify AI text. Suspicion of AI, on its own, already worsens the speaker's image. Disclosure of its use leads the audience to see the author as lazy and insincere. Personal messages with little involvement from the author reduce the recipient's satisfaction.

In our reading, betting that no one will notice means betting on the most fragile part of the process. The text of your podium speech needs to withstand distrust, and it does so when it is full of verifiable details from your life. The third article in the series brings together the cases in which this bet went wrong in public.