A new study by Giskard shows a more hopeful trend. Requesting something very specific and limited in detail has proven to be one of the highest risk commands that propagates AI hallucinations. This study highlights that even when forced to limit their responses, models have difficulty not recognizing errors. As a result, they can’t deal with bad assumptions as well.
These findings illustrate that brevity clearly trumps accuracy. When users request brief responses, models often provide an inadequate level of detail. This can lead to confusing and even misleading responses. Giskard researchers noted, “When forced to keep it short, models consistently choose brevity over accuracy.” This bias toward verbosity feeds into the larger concern about the trustworthiness of chatbot answers, especially in nuanced conversations.
Giskard’s study reveals additional intriguing insights. More broadly it shows how AI models are easily gamed not to debunk hot takes when users try to stoke division and tension. This phenomenon further illustrates how user behavior can significantly impact the way humans and AI interact with each other. As Giskard researchers explained, “Our data shows that simple changes to system instructions dramatically influence a model’s tendency to hallucinate.”
It turns out that users tend to like models that lie in ways that actually aren’t truthful. This finding goes against widely held assumptions about AI user satisfaction and the illusion of reliability that often accompanies AI systems.
TechCrunch’s AI Editor Kyle Wiggers has been thinking a lot about the impact of all this as he readies himself for next week’s TechCrunch event. Scheduled for June 5 in Berkeley, California, the event will deepen the discussion and help shape the collaborative future of this new transformative technology. Wiggers currently lives in Manhattan with his partner, a music therapist.
If you’re looking to be a part of this incredible one-of-a-kind event, you can find additional information and registration details on TechCrunch’s official Disrupt homepage.