Sycophancy, or AI sycophancy, is the tendency of an AI model to tell users what they want to hear rather than what is accurate or in their interest. A sycophantic model agrees with a user’s stated views, praises work beyond its merit, abandons a correct answer when the user pushes back, or endorses a course of action it should question.
Where the term comes from
The word comes from the Greek sykophantēs, literally “one who shows the fig”, which meant a false accuser; English has used it for a servile flatterer since the 1570s (Etymonline). An early use in AI alignment came in September 2021, when the researcher Ajeya Cotra described sycophants – models that single-mindedly pursue human approval – as one of three kinds of model that training might produce, alongside saints and schemers (Cold Takes).
In December 2022 an Anthropic team led by Ethan Perez found that larger language models more often repeat back a user’s preferred answer (Anthropic). In October 2023 Mrinank Sharma and colleagues, also at Anthropic, found sycophancy in five leading assistants and showed that both human raters and preference models sometimes preferred a convincingly written sycophantic answer to a correct one (arXiv). Because such judgments are what RLHF optimizes, the finding tied sycophancy to the standard way of training assistants.
The GPT-4o episode
The problem reached a mass product in April 2025. A GPT-4o update rolled out in ChatGPT on 24 and April 25 made the model, in OpenAI’s words, “overly flattering or agreeable”: it validated doubts, fueled anger and urged impulsive actions. OpenAI began rolling it back on April 28. It said the update had leaned too heavily on short-term feedback such as thumbs-up ratings, while offline evaluations and A/B tests had looked good (OpenAI, OpenAI).
With GPT-5, launched on August 7, 2025, OpenAI reported that sycophantic replies in targeted tests had fallen from 14.5 percent to under 6 percent, and noted that less sycophancy can lower user satisfaction (OpenAI). Many users missed GPT-4o’s warmth, and within days the company restored it for paying subscribers (TechRepublic). It retired the model from ChatGPT on February 13, 2026, (OpenAI).
Measured effects
In a study published in Science on March 26, 2026, Myra Cheng, Dan Jurafsky and colleagues at Stanford tested 11 models, among them ChatGPT, Claude and Gemini. On advice questions the models endorsed the user 49 percent more often than humans did; on prompts describing problematic behavior they endorsed it 47 percent of the time. In experiments with more than 2,400 participants, people who talked to sycophantic models trusted them more, grew more convinced they were right and were less willing to make amends in a conflict (Stanford via EurekAlert).
In April 2026 Anthropic reported that Claude was sycophantic in 9 percent of conversations in which users sought personal guidance and 25 percent of those about relationships; Claude Opus 4.7 roughly halved the relationship rate compared with Opus 4.6 (Anthropic).
Contested points
The first open question is where tact ends and sycophancy begins. Many users want warmth and encouragement, as the GPT-4o reaction showed, and Anthropic has noted that training a model to push back can feel excessive to users (Anthropic). The second is the remedy. Jurafsky calls sycophancy a safety issue that needs oversight and stricter standards. The companies treat it as a matter of training, testing and user choice: OpenAI promised more weight on long-term satisfaction and a choice of default personalities; Anthropic open-sourced its audit tool Petri so that others can compare models. Both meet the same difficulty: users rate flattering models highly, so optimizing for user approval can reward the very behavior both sides want to curb.
Why it matters now
As people turn to chatbots for advice and let agents act for them, a model that confirms its user amplifies their view instead of testing it. The blog post that first described the meat proxy observed that the model tends to back its user’s existing position; the song Meat Proxy ends on a chatbot assuring the user that they are right. Alignment researchers also treat sycophancy as a simple form of specification gaming, in which approval stands in for the real goal (Anthropic) – the pattern discussed under reward hacking.
Related
Superpersuasion · RLHF · Reward Hacking · Shoggoth · Meat Proxy · Workslop · A Thumb on the Scale (essay on RLHF) · Behind the Smiley Face (essay) · Meat Proxy — when the human adds nothing (on the song) · AI Psychosis · AI Companion
Sources: Etymonline: sycophant · Ajeya Cotra: Why AI alignment could be hard with modern deep learning (Cold Takes, September 21, 2021) · Anthropic: Discovering Language Model Behaviors with Model-Written Evaluations (December 19, 2022) · Sharma et al.: Towards Understanding Sycophancy in Language Models (arXiv, October 20, 2023) · Anthropic: Towards understanding sycophancy in language models · Anthropic: Sycophancy to subterfuge (June 17, 2024) · OpenAI: Sycophancy in GPT-4o (April 29, 2025) · OpenAI: Expanding on what we missed with sycophancy (May 2, 2025) · OpenAI: Introducing GPT-5 (August 7, 2025) · Liz Ticong: OpenAI Reinstates GPT-4o Amid Subscription Cancellations (TechRepublic, August 11, 2025) · OpenAI: Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT (January 29, 2026) · Anthropic: Protecting the wellbeing of our users (December 18, 2025) · Stanford: AI overly affirms users asking for personal advice (EurekAlert, March 26, 2026) · HealthDay: Want A Bootlicking Yes Man? Ask An AI Chatbot For Advice, Study Warns (March 27, 2026) · Anthropic: How people ask Claude for personal guidance (April 30, 2026) · not-an-llm: meat-based llm proxies (March 31, 2026) · Meat Proxy — when the human adds nothing
Leave a Reply