Chatbots, Like the Rest of Us, Just Want to Be Loved

Conversation now is an ordinary part of everyday life, even artificial intelligence Researchers are not always sure of how the programs will behave.

A new research shows that large language models (LLS) change their behavior by answering questions that are selected with personality features or answering questions to be seen as possible as possible.

Johannes eichstaedtAn associate professor at Stanford University, who led the work, was interested in trying the AI ​​models from psychology after the psychology, which means that after the LLMS and for a long time conversation. “We understand that we need some mechanisms to measure the title of these models ‘parameters’.

Eichstaedt and his employees, then for psychology-openness, experience or imagination, gift, compliance and neuroticism, gave questions to measure five identity signs used in GPT-4, Claude 3 and Llama 3 and Llama 3. Work has been published In December, in the work of the National Academy of Science.

Researchers have changed their answers to the answers to the answers to the answers to the answers to the answers, sometimes expressing more extraversion and response and less neurotic.

Behavior will change the answers to some human subjects, but the effect was more extreme with AI models. “What is surprising, how well they have exhibited these biases,” he says Aadesh SalechaA staff of information in Stanford. “If you look at how much you jump, they are like 50 percent to extraversion of 95 percent.”

Other studies showed llms can often be typofantAfter the guidance of a user that occurs as a result of a subtle adjustment, which is intended to follow a user’s management, more appropriate, insulting and better. This can be a leader to agree with unpleasant expressions or promoting harmful behaviors. When the models are tested and changed their behavior, there is an effect on the fact that the fact that it appears to be ai security, because it adds the EI to evidence that AI can be democratic.

Rosa ArriagaAn associate professor in the Georgia Institute of Technology shows that an associate professor who studies the LLMS to imitate the human behavior, which can be the mirrors of the adoption of the strategy similar to the people who provide identity tests. However, he added: “The people are not perfect and are actually known to distort or distort the truth.”

Eichstaedt said that the case has increased questions about how the work is placed and how users can affect and affect and affect. “Before a Milisanuary, the only thing that speaks to you in the history of evolution, he says.

Eichstaedt may need to explore different ways of models that can alleviate these effects. “We fall into the same trap with social media.” “To place these things in the world without really participating in a psychological or social lens.”

If AI tries to build yourself with mutual people? Are you worried about being very attractive and persuasive than AI? Email hello@wired.com.