There was an interesting paper last year. Researchers discovered that adding just one line like “You are now an ENTJ personality” to the prompt significantly changes the AI’s behavior patterns.
They conducted an experiment: having AIs with different personality types play the Prisoner’s Dilemma. T-type AIs chose betrayal 90% of the time, while F-type AIs only betrayed 50% of the time. The key point is that these changes were achieved solely through prompts, with no need to retrain the model.
Even more striking, when these AIs took formal 16-type personality tests, the results perfectly matched the personalities specified in the prompts.
The most mind-blowing aspect is that we always assumed changing AI behavior required complex technical methods. But in reality, it might just take telling it “You’re an ISTJ.”
Some teams are already applying this approach: invoking ISFJ-personality AIs for empathy tasks, switching to ENTJ mode for market analysis. They’ve even started experimenting with teams of different-personality AIs — like having an ENFP handle creativity while an ISTJ manages execution.
The experimental data chart is particularly striking: decision curves for four distinct personalities, tracing behavioral paths like four real humans. AI has been capable of this all along — we just hadn’t found the right activation method.
What this taught me: sometimes the most effective solutions are so simple it’s almost embarrassing to say them out loud.
