Listen: Gemini writes better answers when it knows the user is a recognised expert

A recent study by DEJAN reveals that simply changing the user's name in an artificial intelligence system's instructions can dramatically improve the quality of its responses. DEJAN tested Google's latest Gemini model on fifteen expert tasks, asking each question fifteen times. The only variable they changed was the user's name. They used either a recognized authority in that field, a random invented name, or no name at all. The results, scored blindly by a separate AI rater, were striking. When Gemini believed it was writing for a real-world expert, the answers averaged nearly eight out of ten. For a random name, that score dropped to under seven, and with no name, it fell to under six. The expert persona won in twelve out of the fifteen topics. But how did the behavior actually change? For a marketing prompt, instead of outputting generic templates and business acronyms, Gemini delivered highly original, opinionated essays when it thought it was addressing marketing expert Seth Godin. Furthermore, the AI actually thought longer. It used significantly more internal reasoning steps, or thinking tokens, for the experts. It is as if the presence of a prestigious name signals the model to try harder, dig deeper, and bypass standard, cookie-cutter answers. While random names did score slightly higher in three of the fifteen categories, the expert persona consistently outperformed the baseline across the board. This study shows that identity is a powerful lever. By simply framing a query from the perspective of a recognized authority, you can unlock much deeper and more sophisticated reasoning from AI.