Ask Kalicube

The Colleague Fallacy is the false assumption that AI assistants possess associative memory like human colleagues. Because AI converses fluently and references past context, users treat it as if it has a colleague's holistic understanding of shared history. In reality, AI retrieves information through serial keyword search across fragmented knowledge sources - it greps, it does not remember. This mismatch causes users to provide ambiguous, context-poor inputs that would work perfectly with a human colleague but produce poor or wrong results from an AI system that has to search for every piece of context the user assumes is already known.

Research on AI anthropomorphism from Nass and Reeves (The Media Equation, 1996) through Sherry Turkle's work at MIT has documented how humans attribute human qualities to technology. Jason Barnard's Colleague Fallacy (2026) narrows this broad phenomenon to a specific operational failure: users treat AI as if it has associative memory like a human colleague, when it actually retrieves information through serial keyword search across fragmented sources. The distinction matters because it is not about emotional attachment but about working collaboration. Where Gary Marcus has documented AI limitations from a cognitive science perspective, the Colleague Fallacy identifies the specific user behavior those limitations exploit: progressively lazier inputs driven by the illusion of shared context. The concept pairs with The Confidence Fallacy (the output-side problem) to form a complete diagnostic: users give poor inputs because they assume shared context, and they trust poor outputs because the delivery sounds confident. Both trace back to Jason Barnard's Empathy for the Devil (2015): understanding the system's actual mechanisms prevents the failures that anthropomorphism creates.