It seems that when you request an AI to assume the role of an expert, its reliability decreases.

      Requesting AI to act like an expert can have unintended consequences, but researchers may have discovered a remedy.

      You may have come across the suggestion that asking AI to behave like an expert in a particular area results in improved responses. It's widely recommended advice, and it does yield better outcomes on occasion. However, a recent study indicates that employing AI personas might not be as effective as previously believed.

      Researchers from the University of California assessed 12 different personas across six language models, ranging from experts in math and coding to creative writers and safety monitors, aiming to evaluate AI's performance when directed to operate as an expert.

      The findings were varied. While using a persona made the AI appear more professional and allowed it to adhere to guidelines more effectively, it also hindered its ability to recall facts accurately. The study suggests that invoking an AI persona transitions it into a mode focused on following instructions rather than retrieving knowledge, leading to a compromise in accuracy.

      What’s the remedy?

      To address this issue, the researchers created PRISM, which stands for Persona Routing via Intent-based Self-Modeling. Rather than always utilizing a persona or never using one, PRISM enables AI to determine the best approach for itself.

      When posed with a question, PRISM generates two responses: one from its standard mode and another from its persona. It then evaluates both and provides the answer that is more effective for the given query.

      The expert response is not eliminated even if the standard answer prevails. Instead, the reasoning style is retained in a lightweight component named a LoRA adapter, which the AI can access later when necessary. This solution appears straightforward, yet proves to be efficient.

      How did PRISM perform?

      PRISM improved AI’s overall score by one to two points on the MT-Bench, a test that assesses how well an AI adheres to instructions and remains useful. For tasks involving writing and safety, the personas were beneficial. However, for straightforward knowledge questions, forgoing the persona was the superior choice.

      The researchers intend to evaluate PRISM with additional personas and enhance its capacity to deliver improved answers. Although this is still in the early stages, it has the potential to transform how we interact with AI in the future.