AI creates an image of your personality – in just two hours

Researchers at Stanford University and Google subsidiary DeepMind have developed an AI that can portray the personality of real people. A two-hour interview should be sufficient.

Imagine having a two-hour conversation with an AI. This then creates a digital image of your personality that reflects your values ​​and preferences with astonishing accuracy. This idea does not come from a science fiction film, but is reality.

Researchers at Stanford University and Google subsidiary DeepMind have developed so-called simulation agents using artificial intelligence. These should be able to imitate human behavior with an accuracy of up to 85 percent.

Researchers create AI images of human personalities

To create their simulation agents, the researchers prepared special interviews. The test subjects had to talk about personal experiences, beliefs and thoughts on social issues. The information was then fed into an AI model that generated an image of the person in question.

The aim was to simulate the participants' decisions and reactions. The digital personalities generated open up new opportunities for research. Social scientists could use them, for example, to carry out complex studies that would be too expensive or unethical with real subjects. Companies could use simulation agents to test behaviors or market reactions.

High potential for abuse

But there are also dark sides. The technology could be misused to digitally replicate people without their consent. This poses the risk that such digital images may say or do things that real people would never have intended. The accuracy of the simulations is also not perfect, especially for tests that examine ethical values ​​such as fairness.

Nevertheless, the study shows that digital personalities can be created with surprisingly little data – such as a two-hour interview. Companies like Tavus, which already offer “digital twins”, see great potential in this and want to further develop the method.

Also interesting: