Study Finds Most LLMs Accept False Code Claims, Even With Supporting Context

A developer tested 14 large language models against 50 facts drawn from a 50,000-line Python codebase to measure how often models incorrectly validate false memory claims. The experiment ran two conditions per fact: one where models saw only the claim, and one where they also received code context and supporting patterns. Several models, including nemotron-3-nano and glm-4.7-flash, accepted nearly one in four to one in three false claims even when given supporting code anchors. Top-performing budget models from the Qwen3 family matched Claude's false-accept rate of zero at a fraction of the cost, though all models universally accepted one specific false claim tied to a misleading keyword anchor. The findings suggest that cheap models are not universally reliable for memory verification and that anchor-based prompting can itself introduce contamination risk.