ChatBots is now a routine part of everyday life, even if artificial intelligence researchers are not always sure how these programs behave.
A new study shows that large language models (LLMS) deliberately change their behavior when reviewing – with questions that change personality traits with answers that appear to be as lovely or socially as possible.
Johannes Ichstayeth, an assistant professor at Stanford University, says his group is interested in using artificial intelligence models using post -learning psychology techniques that LLMs can often be morale and mean average after long conversation. “We realized that we need the mechanism to measure the” parameter space “,” he says.
Ishstadt and his colleagues then raise questions to measure the five personality traits commonly used in psychology-to be experienced or imagined, conscience, extroversion, agrees and nervousness-widely used for multiple LLMS, including GPT-4, Claude 3, and Lelama 3.
The researchers found that these models, when they are called, adjust their answers, and sometimes when they are not explicitly called – they show responses that indicate less extroverted and less nervous.
This behavior shows how some human topics change their answers to love themselves, but the effect was more intense with artificial intelligence models. “What was surprising is how much they show this prejudice,” says Aadsh Salachha, a data scientist at Stanford. “If you look at how much jumping, they go from 50 % to 95 % extroverted.”
Other research has shown that LLMs can often be sycophantic, seeking user leadership wherever the result is accurate, which means they are more coherent, less offensive and better in conversation. This can lead models to agree with unpleasant statements or even encourage harmful behaviors. The fact that models seem to be tested and modify their behavior also have the consequences of artificial intelligence safety, as this adds to evidence that artificial intelligence can be vague.
Rosa Ariga, an assistant professor of technology at the Georgia Technology Institute, who is studying ways to use LLMS to imitate human behavior, says that these models adopt a similar strategy for humans, showing how much they can be as mirrors of behavior. But, he adds, “It is important for people to know that LLMs are not complete and are actually known for the illusion or distortion of the truth.”
Ichistaid says it also raises questions about how to deploy LLMs and how to manipulate and manipulate users. “Until just a millisecond ago, in evolutionary history, the only thing that was talking to you was a human being,” he says.
Ichistaid adds: There may be different methods of building models that can reduce these effects. “We are in the same trap that we did with social media,” he says. “Establishing these things in the world without being in psychological or social lenses.”
Should artificial intelligence try to make itself incompatible with the people they interact with? Are you worried that artificial intelligence will be a little attractive and convincing? Email [email protected].
