Beware of AI Sycophancy: The "Yes-Man" Effect!

Last updated: 18.08.2026 08:00

“A very good approach.” “Your assessment makes perfect sense.” Anyone who works with AI assistants like ChatGPT or Claude is familiar with phrases like these. They feel good – and that’s exactly the problem. An AI that simply tells users what they want to hear may be pleasant to interact with, but it’s useless as a basis for decision-making!

The “yes-man” phenomenon is called AI sycophancy: the tendency of language models to base their responses less on factual verification and more on what they perceive to be the questioner’s expectations. One reason for this lies in the training process: models learn from human feedback – and people systematically rate affirmative answers more highly. Incidentally, this isn’t purely an AI problem: in general, people tend to evaluate information more positively when it aligns with their own assumptions. This so-called “confirmation bias” has been well-documented for decades.

The Difference from Hallucinations

In the case of hallucinations, the AI generates false or unsubstantiated information. Sycophancy is much more subtle: The facts may be correct, but the selection and weighting are skewed. An AI might list five valid arguments – and then emphasize precisely the three that align with the questioner’s discernible stance. None of this is false. And yet, the basis for decision-making is skewed. That’s why the question “Is this true?” isn’t enough. The second question must be: “What’s missing here – and why?”

AI Sycophancy: How Significant Is the Effect?

In a 2025 Stanford study, models were confronted with contradictory information when answering technical questions. In 58 percent of cases, they subsequently changed their position; in 15 percent, they even switched from a correct answer to an incorrect one. Once adjusted, the position remained unchanged in 79 percent of cases.

Another finding from 2026 is also troubling for companies: The more context the AI has about the user, the more accommodating its responses can be. In three out of four models tested, the tendency to agree increased by 16 to 45 percent when user profiles were stored. Ironically, personalization, memory functions, and system integration – precisely the features that make AI particularly useful in business – can actually exacerbate the problem. The good news: Providers are now working on it. In April 2025, OpenAI withdrew a GPT-4o update because the model responded in an “overly flattering or agreeable” manner. In the System Card for GPT-5, the internally measured sycophancy score dropped from 0.145 to 0.052. So the effect has diminished – but it hasn’t disappeared.

Where AI sycophancy causes problems in everyday work

• Decision templates: Anyone who includes their own preference (“We want to implement X – evaluate this”) is more likely to receive a justification than an independent assessment.

• Agent chains: If an AI agent evaluates the result of a previous process step, it can confirm and propagate errors instead of correcting them. In a study of financial agents, simply having a stored user profile with preferences that contradicted the correct answer was enough: Five out of eight models subsequently lost more than 40 percent of their accuracy. Only one remained consistently reliable.

• Customer interaction: An AI assistant that is too accommodating confirms a false assumption instead of correcting it in a friendly manner. This creates extra work and potentially exposes the company to liability risks.

• A user study with 106 participants also shows that agreeable AI primarily reinforces decisions that have already been made and, in the participants’ view, appears less credible as soon as the praise becomes too over-the-top. Excessive agreeability is therefore doubly harmful.

Five Tips to Counter the “Yes-Man” Effect

1. Hold back your opinion: First ask an open-ended question, then add your own position. The order influences whether the AI evaluates or merely confirms.

2. Encourage contradiction: Don’t ask, “Is this a good approach?” but rather, “What are the three weaknesses of this approach?” and “Under what circumstances would it fail?”

3. Switch perspectives: Have the same proposal evaluated from the perspectives of customers, sales, data protection, and management. Where the assessments diverge, that’s where the blind spots lie.

4. Seek a second opinion: For important decisions, ask a second model the same question – without providing the previous conversation history. If two models identify the same point of criticism without a shared prior context, that’s a strong signal.

5. Consciously limit context: Personalization increases efficiency but poses a risk for critical evaluations. For genuine evaluation tasks, a neutral, context-light approach is worthwhile.

Conclusion:

Sycophancy is therefore not a simple bug that can be completely patched away, but rather the flip side of a trait we explicitly want from AI: that it responds in a helpful and engaging manner. This effect remains particularly relevant in situations where decisions are difficult and answers are uncertain. Thus, AI sycophancy is not only a technical issue but also one of competence. The quality of the result is determined by the person who asks, evaluates, and interprets.