MENU
A person wants to use an AI model to predict the winner of an athletic event. The person repeatedly prompts the model until it chooses the person's favorite athlete as the winner. What is the type of bias described in the scenario?
This scenario is a textbook example of Confirmation bias. Unlike other biases that reside within the data or the algorithm, confirmation bias is a cognitive bias on the part of the user. It occurs when a person searches for, interprets, or prioritizes information in a way that confirms their pre-existing beliefs or desires. By repeatedly prompting the AI until it provides the 'desired' answer, the user is disregarding all previous outputs that contradicted their preference.
In the context of prompt engineering, confirmation bias can lead to 'leading prompts' where the user subconsciously (or consciously) steers the AI toward a specific conclusion (e.g., 'Tell me why Athlete X is the best'). This undermines the AI's value as an objective tool for analysis. To mitigate this, prompt engineers should practice 'neutral prompting' and seek to explore multiple perspectives (using techniques like Tree of Thought) rather than hunting for a specific output. Failing to recognize confirmation bias can lead to poor decision-making and the creation of 'echo chambers' where AI is used to justify subjective opinions rather than uncover objective truths.