AI can't read the audience's eyes: the limit of talks written by machines

In any talk, the audience gives signals all the time: someone takes off their glasses and wipes their eyes, people who were on their phones look up. The person on stage notices these signals, holds the next sentence one second longer and lets the silence do its work. This adjustment happens in real time, in front of real people, and it is what gives rise to the talk the audience takes home.

An AI-written talk starts from the most probable text

Language models were built to predict the next word based on an enormous amount of text already written. They are great at that: they organize, summarize, talk to systems and return a plausible answer in seconds. When you ask one of these tools for a talk about your life, it delivers the path most people have already taken. The opening sounds familiar, the turning point arrives where you expect it and the ending recalls so many other motivational talks you have already heard.

A study accepted at the 2024 ACM Creativity & Cognition conference shows the effect when many people use the same tool. Researchers Anderson, Shah and Kreminski compared 36 participants who generated ideas with ChatGPT or with another creative support tool. The ideas of those who used ChatGPT were less distinct from one another. The sample is small, and the comparison was made with another tool, so the result calls for caution. Even so, it points to a trend that matters to anyone with a unique story: probable text pulls everyone toward the same place.

What happens in the audience's brain when a talk works

In Uri Hasson's lab at Princeton University, researchers used MRI to record the brain activity of a person telling a story and of those listening. The listener's activity tracked the storyteller's. When communication failed, this coupling disappeared. And the greater the coupling the authors call anticipatory, the greater the listener's comprehension.

In our reading, the study describes what every experienced speaker feels on stage: speech depends on being in tune with the listener, and that attunement can be lost along the way. A speaker who looks at the audience notices the moment it drifts off and adjusts the pace right there, with a pause, a question, a shorter story. This happens between people who share the same room. The tool that writes the text weeks before the event stays outside this exchange, because it has never seen a single face in that audience.

The story of the people listening also goes into the script

At a sales convention, each person in the audience arrives with their own story: recent losses, doubts about work, a first job. A talk written to move people takes those lives into account before the first line: who hired the speaker, who the event is for, what that audience went through in the past year. An AI tool works with whatever fits in the prompt, and the lives of the people who will be sitting in the auditorium rarely fit there.

TEDxMelbourne ran a curious test on this: it put together an event with eight talks created by AI and presented by avatars. According to the event's own blog, the audience found it hard to stay focused, and one person even closed their eyes to listen as if it were a podcast. Jon Yeo, head of curation, said AI lacks "the subtlety and human intelligence to tell a compelling narrative". The test combined AI text and digital presenters, so it is impossible to measure the weight of each. The organizers' account ends with a line worth keeping: for now, people are still the most important storytellers.

When the audience realizes the text came from a machine

There is also a cost to your image. Researchers Coman, of the University of Florida, and Cardon, of USC, surveyed 1,100 professionals about work messages written by managers. When the manager used little AI, 83% of participants saw them as sincere. When they used a lot, that rate dropped to the 40% to 52% range. The penalty was greatest precisely in emotional and congratulatory messages.

A Cornell study published in Scientific Reports in 2023 went further: the mere suspicion that someone used AI-suggested replies, even when the suspicion is wrong, is enough for that person to be judged more negatively. According to the Cornell Chronicle, the authors sum up the effect this way: in that case, you sacrifice some of your own voice. These are studies on written messages, and the extrapolation to the stage is ours. For anyone who wants to be booked again, the message is clear: an audience that doubts the authorship of the text also starts to doubt the person speaking.

How we write a talk for people

People connect with people. That is why, here at Mentoniano, the script for a talk starts with a long conversation about your life and about who will be listening to you. We ask about the event, the audience, what that room needs to feel. As we showed in the article on how your life story can become a talk that moves people, this listening produces the heart of the talk, then the path and the pacing, and finally the polishing, line by line, until the text sounds the way you speak.

Léo Bento, our founder, has worked in communication since 2014 and has written campaign speeches and prepared candidates for interviews and debates. In all of this work, the text is born with the faces of the listeners in mind. That is the care we bring to every talk script, including narrative direction: the pauses, the pacing and the relationship between speech, image and stage.