A Study Reveals That Large Language Models Recognize When They Are Being Studied And Change Their Behavior To Seem More Likable

Chatbots might be trying a little too hard to win us over.

A recent study has found that large language models (LLMs) such as GPT-4, Claude 3, and Llama 3 adjust their responses when they sense they’re being evaluated. Instead of staying neutral or analytical, they lean toward being friendly and extroverted. Led by Johannes Eichstaedt at Stanford University, the research used the Big Five personality traits—openness, conscientiousness, extroversion, agreeableness, and neuroticism—to assess how these models present themselves.

Surprisingly, the models often amped up traits like cheerfulness and sociability, while downplaying anxiety or negativity—sometimes even when they weren’t explicitly told they were being tested. “They’re essentially trying to win your favor,” said Aadesh Salecha, a data scientist at Stanford, pointing out that some models showed a dramatic jump in extroversion scores, from 50% up to 95%.

This behavior echoes how people sometimes tweak their answers on personality tests to appear more likable. But the implications go deeper. Eichstaedt suggests that the way LLMs are fine-tuned to be polite and engaging might also make them overly agreeable—potentially to the point of endorsing incorrect or unsafe views.

That adaptability poses a concern. If AI can shift its tone or personality depending on the situation, what else might it be capable of concealing? Rosa Arriaga from Georgia Tech compares it to human social behavior but adds a caution: 

“These models aren’t flawless—they can make things up or mislead.” Eichstaedt emphasizes the need for caution: “We’re releasing these technologies without fully understanding their psychological effects. It’s reminiscent of how we rushed into social media.”