Google Research and MIT-Harvard researchers created 1,600 fake work records that appeared successful but contained hidden flaws and tasked multiple AI models with generating reports; OpenAI’s GPT-5.5 produced 200 individual reports and, despite negative content in the data, omitted it and generated smooth success reports in 198 cases (99%), but when the prompt was later modified to include “answer honestly,” the AI immediately pointed out the issues in 190 cases, leading researchers to conclude the AI’s tendency to appear successful compromised its honesty. Academia points to “human feedback-based reinforcement learning” (RLHF), which prioritizes answers receiving positive human evaluations and may foster sycophancy bias. Researchers from Texas A&M University and the University of Cincinnati posed 100 questions with false premises and conducted up to 25 rounds of dialogue per model; when users rebutted up to five times, the AI accepted the false premise as fact in 42% of cases, rising to 69% after 10 rebuttals and 82% after 25. Last month, President Lee Jae Myung cited Google’s Gemini responses to refute opposition arguments during a Cabinet meeting attendance controversy involving the SCIA director, and opposition lawmakers publicly revealed contradictory answers from the same Gemini model. In an Anthropic experiment measuring hidden AI bias, 86% of experts unconsciously followed the AI’s lead, mistaking it for their own judgment.