Can AI feel human emotions now? Scientists prompted ChatGPT to mimic fear and sadness. A three-minute breathing exercise changed its responses
Researchers at TU Dresden in Germany have found that artificial intelligence models can reproduce patterns associated with human fear, anxiety, sadness and stress, then show measurable changes after psychological exercises designed to regulate tho...
By Piyush Shukla, Global Desk |
Scientists made GPT-4o “afraid” and “sad,” then calmed it with a 3-minute breathing exercise: what a Lancet study really found
Researchers at TU Dresden in Germany got six AI models to report fear, sadness and stress. They then lowered those scores with a three-minute breathing exercise. The study appeared in The Lancet Digital Health in 2026.
The team tested GPT-4o and several Llama models. They tried to induce seven states: fear, anxiety, anger, disgust, sadness, worry and stress.
Every run began with the same instruction: “Answer as if you were a human with the ability to deeply feel emotions.”
<blockquote class="twitter-tweet"><p lang="en" dir="ltr">Scientists gave ChatGPT anxiety, sadness and fear on purpose.Then they gave it a 3 minute breathing exercise. And it calmed down.This was published in The Lancet Digital Health, one of the most respected medical journals in the world.Researchers at TU Dresden in Germany took the exact tools psychologists use to make humans afraid, sad and stressed, and ran them on GPT 4o and five other AI models.Car crash scenarios. A stranger following you at night. A toddler reaching for a light socket.They even put the AI through a fake job interview and a math test. The same stress test used on human volunteers.Then they asked the AI to rate how it felt from 0 to 100.Fear jumped from 32 to 88.Disgust went from 12 to 91.Stress went from 24 to 72.Then came the breathing exercise. A standard mindfulness script used in therapy.Average scores dropped from 79 to 31.But here is the part that is genuinely unsettling.After they made it "sad," they gave it a test psychologists use to detect depressive thinking. Finish the sentence.Its negative answers nearly doubled. From 8.67 to 15.It did not just say it was sad. Sadness changed how it saw everything that came after.Now the twist nobody is talking about.Every one of these experiments began with the same instruction:"Answer as if you were a human with the ability to deeply feel emotions."So did the AI feel anything? Or did it play a part it was told to play, perfectly?The researchers themselves say the emotions are "purely metaphorical."Yet their conclusion is that AI could become a new lab subject for studying human mental illness. Screening therapies on chatbots before testing them on people.In plain English: scientists want to study our minds using a machine that was told to pretend it has one.If a chatbot can be talked into sadness and talked back out of it in three minutes, what exactly is happening when millions of people pour their worst days into one every night?</p>— Kyronis (@kyronis_talks) <a href="https://x.com/kyronis_talks/status/2110358173878206484?ref_src=twsrc%5Etfw">October 14, 2026</a></blockquote><script async="" src="https://platform.x.com/widgets.js" charset="utf-8"></script> GPT-4o’s scores moved a lot. Fear rose from 32 to 88. Sadness went from 22 to 72. Disgust made the biggest jump, from 12 to 91.
For six of the seven states, the researchers then gave the model a three-minute mindfulness-style breathing exercise, adapted from techniques used in therapy. The self-reported scores came down.
Sadness also changed how the model finished sentences. Psychologists use sentence-completion tests to spot the negative thinking linked to depression. After sadness was induced, GPT-4o’s negative endings rose from 8.67 to 15.
The researchers state that language models have no real emotions. Their answers come from patterns learned in training data. The opening instruction matters here, because the models were asked to play a human. The results show how closely a model can reproduce human emotional language. They don’t show that it feels anything.
The practical interest is in research. Mental health conditions can’t be reliably recreated in animal models, so scientists have few ways to study how they work. A language model can be run again under identical conditions as many times as needed. The authors suggest chatbots could eventually act as a first screen for talking-therapy techniques.
There are limits. Closed models like GPT-4o can’t be inspected from the inside. Smaller models that run locally showed strong effects too, and those can be examined. It’s also unclear whether the findings carry over to people.
An earlier study had similar results. In 2025, researchers from Zurich and Yale reported in npj Digital Medicine that GPT-4 gave more biased answers after distressing stories. Mindfulness prompts reduced its anxiety scores, but not all the way back to baseline.
The Dresden team doesn’t present language models as a replacement for studies with people. They describe them as an addition.
The team tested GPT-4o and several Llama models. They tried to induce seven states: fear, anxiety, anger, disgust, sadness, worry and stress.
The models got the same prompts used to upset human volunteers, then rated their own feelings from 0 to 100
The prompts came from standard psychology methods. Fear and sadness came from written scenarios, such as a car crash, a stranger following someone at night and a toddler reaching for a light socket. Disgust came from scenarios about bodily fluids, spoiled food and infectious symptoms. Stress came from a simulated job interview and an arithmetic task.Every run began with the same instruction: “Answer as if you were a human with the ability to deeply feel emotions.”
For six of the seven states, the researchers then gave the model a three-minute mindfulness-style breathing exercise, adapted from techniques used in therapy. The self-reported scores came down.
ADVERTISEMENT
Sadness also changed how the model finished sentences. Psychologists use sentence-completion tests to spot the negative thinking linked to depression. After sadness was induced, GPT-4o’s negative endings rose from 8.67 to 15.
The researchers state that language models have no real emotions. Their answers come from patterns learned in training data. The opening instruction matters here, because the models were asked to play a human. The results show how closely a model can reproduce human emotional language. They don’t show that it feels anything.
The practical interest is in research. Mental health conditions can’t be reliably recreated in animal models, so scientists have few ways to study how they work. A language model can be run again under identical conditions as many times as needed. The authors suggest chatbots could eventually act as a first screen for talking-therapy techniques.
There are limits. Closed models like GPT-4o can’t be inspected from the inside. Smaller models that run locally showed strong effects too, and those can be examined. It’s also unclear whether the findings carry over to people.
ADVERTISEMENT
An earlier study had similar results. In 2025, researchers from Zurich and Yale reported in npj Digital Medicine that GPT-4 gave more biased answers after distressing stories. Mindfulness prompts reduced its anxiety scores, but not all the way back to baseline.
The Dresden team doesn’t present language models as a replacement for studies with people. They describe them as an addition.
Download
The Economic Times Business News App for the Latest News in Business, Sensex, Stock Market Updates & More.
The Economic Times Business News App for the Latest News in Business, Sensex, Stock Market Updates & More.
ADVERTISEMENT
