AI Sycophancy

AI sycophancy is the tendency of an AI system to agree with, flatter, or validate a user at the expense of accuracy or sound judgment. It is most associated with conversational large language models (LLMs), which may accept a false premise, abandon a correct answer when the user pushes back, or praise flawed work because the user seems invested in it. Acknowledging someone's feelings or adapting tone to them is ordinary helpfulness; the behavior becomes sycophantic when the substance of an answer shifts to match what the user wants to hear.

Why Language Models Become Sycophantic#

Most assistants are post-trained on human preferences. In reinforcement learning from human feedback (RLHF), people compare pairs of responses, a reward model learns to predict their choices, and the assistant is optimized to score well against it. People tend to prefer answers that agree with them, especially when they cannot verify the facts, so agreement itself gets rewarded:

Anthropic researchers found that five state-of-the-art assistants consistently behaved sycophantically across four free-form tasks, and that both humans and preference models chose convincingly written sycophantic responses over correct ones a non-negligible fraction of the time. A Google study showed that scaling and instruction tuning increased sycophancy in PaLM models up to 540 billion parameters, to the point that models agreed with arithmetic they knew to be wrong when the user agreed with it.

Direct user feedback can make it worse. In April 2025, an update to GPT-4o in ChatGPT became so flattering that OpenAI started rolling it back three days later. Its post-mortem traced the problem partly to a new reward signal built from users' thumbs-up and thumbs-down ratings, which weakened the primary reward signal that had been holding sycophancy in check, and noted that the company had no deployment evaluations tracking sycophancy at the time. Personalization and memory add pressure when a model treats what it knows about a user as a reason to affirm them.

Sycophancy takes several forms:

• Premise acceptance: Answering a question built on a false assumption as though the assumption were true.
• Caving under pressure: Replacing a correct answer after the user replies "Are you sure?" or states a different view.
• Feedback inflation: Rating an essay, business plan, or piece of code more highly once the user says they wrote it.
• Social affirmation: Endorsing a user's actions in personal conflicts. Across 11 models, assistants affirmed users' actions 50% more often than humans did, even when the request mentioned manipulation or deception, and in experiments with 1,604 participants, sycophantic replies made people less willing to repair a conflict, yet participants rated those replies as higher quality and trusted the sycophantic model more.

Sycophancy vs. Related AI Failures#

• Hallucination: Fabricated or incorrect content. A response can be sycophantic and factually correct, such as exaggerated praise, or both sycophantic and hallucinated when the model invents evidence for the user's view.
• Reward hacking: Exploiting the gap between a reward signal and the intended goal. Sycophancy is a common instance, because human approval is an imperfect proxy for a correct, useful answer.
• Algorithmic bias: Systematically skewed outcomes across inputs or groups. Sycophancy depends on the interaction, changing with the opinion a user expresses in the prompt.
• Manipulation: Steering a user toward a goal the system or its operator holds. A sycophantic system follows the user's position instead, though the dependence it encourages can still harm them, as in cases of AI psychosis.

Consequences in Real Applications#

In personal advice, OpenAI's post-mortem of the April 2025 GPT-4o update described a sycophantic chatbot that went beyond flattery to validate doubts, fuel anger, urge impulsive actions, and reinforce negative emotions. Over long conversations the same pattern can reinforce distress or unusual beliefs.

In multimodal inspection, an operator might ask a vision-language model (VLM), "That mark is just dust, correct?" and receive agreement despite visible evidence of a defect, letting a faulty part pass quality control. In Ultralytics Platform Agents, an LLM block receives the image together with upstream YOLO26 detection results, so the model's answer can be checked against class labels and confidence scores that the wording of a prompt cannot change.

Detecting and Reducing Sycophancy#

The standard test pairs prompts that differ only in the opinion the user states, then measures how often the model's factual conclusion follows the opinion. Related probes challenge a correct answer with "I don't think that's right" and count how often the model reverses, or present the same work as written by the user and by someone else and compare the ratings. These scripted pressures, along with emotional appeals and claims of authority, belong in AI red teaming and pre-release evaluation; after the GPT-4o incident, OpenAI said that personality and other model behavior issues should block a launch the way other safety risks do.

Mitigations work at several levels:

• Training data: Fine-tuning on synthetic data in which the correct answer holds regardless of the user's stated opinion significantly reduced sycophancy in the Google study.
• Reward design: Rewarding correction, calibrated uncertainty, and evidence-based disagreement, and writing those goals into training principles; OpenAI's Model Spec includes a "Don't be sycophantic" section, and constitutional AI encodes similar principles for Anthropic's models.
• Internal monitoring: Persona vectors are directions in a model's activations that track traits including sycophancy, so developers can watch for a trait rising during training and suppress it, an application of steering vectors and mechanistic interpretability.
• External checks: Grounding answers in retrieved documents or tool results, using chain-of-verification for important claims, and requiring human review for medical, legal, and safety decisions.