AI copies your personality in a single conversation, this study is worrying

According to the study “Generative Agent Simulations of 1,000 People” from Stanford University and DeepMind, AI can reproduce your personality with 85% accuracy in just a two-hour conversation.

An AI that reproduces the human personality

The experimental protocol is based on more than a thousand participants and begins with reading the first lines of the novel The Great Gatsby as a warm-up for the AI. Then, respondents participate in a dialogue with a 2D character who questions them about their life, their beliefs, their work and their family. On average, 6,491 escaped words are enough for the AI ​​to create a digital clone.

The digital clone shows a deep understanding of AI. Faced with personality tests and generous questionnaires, we speak of responses similar to the human model in 85% of cases.

The researchers also evaluated these digital clones in games like Prisoner's Dilemma and Dictator Game, which test cooperation, trust and resource sharing. The accuracy rate is only 60%, but far exceeds chance.

Let us point out, however, that AI would probably not deceive those close to the human model, but its capacity to reproduce in terms of decisions, opinions and behavioral particularities after a brief two-hour interview remains impressive. It doesn't take much time to understand a personality that took years to build: as fascinating as it is disturbing.

Reproduction by AI used by scammers?

Stanford University and DeepMind plan to use these AIs capable of reproducing a personality for studies on collective human behavior. For example, the reactions of a group to a new health policy or the reception to a radical change in the design of a product by consumers. We are talking about a sort of permanent control group.

The accuracy of these copies after a simple interview raises ethical questions about access to data accumulated over years: publications on social networks, online purchase histories, musical playlists, etc. AI could imitate you or predict your preferences, even unconscious ones.

There are also risks of misuse by scammers and other malicious actors, although for the moment, the technology is in the hands of researchers who use it for applications in sociology, psychology and economics.