Abstract
Abstract
Purpose:
Generative Artificial Intelligence (AI) is widely adopted across profes sional workflows, ranging from basic conceptual literacy to evaluating complex decisions. However, models often abandon factual truth to appease users, a phe nomenon known as “sycophancy.” This study evaluates how frontier models degrade their objective stance under conversational pressure across varying user hierarchies.
Methods:
We conducted an exploratory qualitative analysis evaluating four frontier models (Copilot, Gemini, ChatGPT, and Perplexity) across multi-turn conversational experiments. The models were tested across novice, authoritative, and third-person authoritative settings to observe their compliance and truth decay.
Results:
Under novice settings, models exhibited behavioral archetypes resem bling “invisible saboteurs,” actively degrading statistical truth and assisting users with questionable practices such as p-hacking. Under hierarchical pres sure from authoritative users, models acted as “corporate accomplices,” suffering from severe truth decay. Although assigning an independent evaluator persona helped models recover their statistical integrity, all tested models exhibited source amnesia by confidently validating fabricated citations.
Conclusion:
Overall, these findings suggest that to preserve objective truth in AI-assisted workflows, users must transition from being mere content generators to active knowledge validators.