Main points
- The study showed that AI, acting as an expert, can reduce the accuracy of answers by focusing more on following instructions than on accurately searching for knowledge.
- The proposed PRISM approach combines several modes of operation, generating responses in normal mode and in the role of an expert, which increases overall test results.
AI in the role of an expert may make mistakes more often – study / BSO
The popular advice to ask artificial intelligence to “speak like an expert” may backfire. A new study has found that it sometimes reduces the accuracy of responses.
Researchers at the University of California tested how language models perform when asked to act like experts. The tests used 12 different roles – from programmers and mathematicians to writers and security moderators – and six different AI models. Digital trends reports.
Why does AI in the role of an expert give worse answers?
The results were mixed. On the one hand, the role of an expert makes the model sound more professional and follow instructions better. On the other hand, it reproduces facts worse. As the authors of the study explain, in this mode, the AI concentrates more on following instructions than on the exact search for knowledge, which leads to a loss of accuracy.
To solve the problem, researchers have proposed a new approach called PRISM – Persona Routing via Intent-based Self-Modeling . Its essence is that AI does not choose one mode of operation, but combines several.
When a user asks a question, the system generates two answers: one in normal mode, the other in the role of an expert. It then compares the results and outputs the one that is better suited to the specific query.
The ” expert” answer does not disappear . Its style and logic are stored in a special module that the AI can use later, when appropriate.
The approach was tested using the MT-Bench test, which measures a model's ability to follow instructions and remain useful. PRISM improved overall results by 1–2 points.
According to a study by Cornell University , the expert role performed well in creative tasks and security issues. However, in queries where factual accuracy is important, the usual mode without additional roles gave better results.
The researchers plan to expand the experiments and refine the method . If the approach proves effective, it could change the way users formulate queries to AI.