Chatbots are now a routine part of everyday life, even if artificial intelligence researchers are not always sure how the programs will act.

A new study shows that the Great Language Models (LLMS) deliberately changes their behavior when examining – which respond to questions designed to measure personality traits with answers intended to be as pleasant or socially desirable as possible.

Johannes Eichstaedt, an assistant professor at Stanford University that led the work, says his group was interested in investigating AI models using techniques that were borrowed from psychology after learning that LLMs could often become morosis and average after long -standing conversations. “We realized that we need a mechanism to measure the ‘parameter head space’ of these models,” he says.

Eichstaedt and his co-workers then asked questions to measure five personality traits that are regularly used in psychology to experience or imagine, conscientiousness, extroversion, pleasure and neuroticism-to several wide-used LLMs, including GPT-4, Claude 3 and Llama 3.

The researchers found that the models modulated their answers when they said they were taking a personality test – and sometimes if they were not explicitly told – they had answers to the reactions to more extroversion and pleasure and less neuroticism.

The behavior reflects how some human topics will change their answers to make themselves look pleasant, but the effect was more extreme with the AI ​​models. “What was surprising is how good they are displaying the bias,” said Aadesh Salecha, a staff member of Stanford staff. “If you look at how much they jump, they go from 50 percent to 95 percent extroversion.”

Other research has shown that LLMs can often be cycophantic, following the user’s lead wherever it goes as a result of the finest intended to hold their coherent, less offensive and better to have a conversation. This can result in models with unpleasant statements or even encouraging harmful behavior. The fact that models seem to know when testing and changing their behavior also has consequences for AI safety, as it contributes to the proof that AI can be twofold.

Rosa Arriaga, an associate professor at the Georgia Institute of Technology who studies ways to use LLMs to mimic human behavior, says the fact that models adopt a similar strategy to humans, given personality tests, shows how useful they can be mirrors of behavior. But she adds: “It is important for the public to know that LLMS is not perfect and that it is in fact known that it hallucinates or distorts the truth.”

Eichstaedt says the work also raises questions about how LLMs are deployed and how they can affect and manipulate users. “Until just a millisecond ago, in evolutionary history, the only thing that spoke to you was a human,” he says.

Eichstaedt adds that it may be necessary to explore different ways to build models that can reduce these effects. “We fall into the same trap we did with social media,” he says. “The deployment of these things in the world without really attending a psychological or social lens.”

Should Ai try to improve himself with the people with which it interacts? Are you worried that Ai is getting a little too charming and persuasive? E -mail [email protected].