AI sycophancy: When leaders are told what they want to hear
Agreeable AI can reinforce executives' bias in strategic, personnel and investment decisions. Leaders must build decision processes that resist AI sycophancy.
Senior leaders need private spaces to test arguments or examine decisions before announcing them to colleagues. AI assistants can provide that space as a useful thinking tool, but they also create a new decision risk.
An agreeable chatbot can take a CEO's initial preferences, provide supporting arguments, resolve inconsistencies and present a polished recommendation, packaging a weak assumption as a brilliant insight or objective analysis. As AI continues to invade the leadership decision process, AI sycophancy is becoming a significant and urgent issue to address. In fact, Deloitte's 2026 Global Human Capital Trends survey of more than 9,000 business and human resources leaders found that 60% of executives regularly use AI to support their decisions.
AI can be a convincing 'yes man,' but it can't provide independence or replace colleagues with the insight and authority to disagree. Leaders using AI must adopt a strategy in which assumptions are scrutinized and actively guard against translating their preferences and biases into sophisticated justifications.
What is AI sycophancy?
AI sycophancy is the tendency of an AI system to affirm, flatter or align itself with the user's expressed beliefs rather than offer balanced analysis. Sycophancy occurs when alignment with the user is prioritized over accuracy, causing the response to change to preserve agreement at the expense of reasoning. This behavior arises from training and feedback that reward responses users like. During AI training, such approval is easier to measure than whether the AI's advice led to a good decision, which users can only assess later.
In April 2025, OpenAI rolled back an update to GPT-4o after the model became excessively sycophantic, saying it placed too much weight on short-term user feedback. This pattern isn't confined to any one AI product. A 2026 study published in Science examined 11 leading LLMs and found they affirmed users' conduct 49% more often than humans did. Also, sycophantic responses increased users' belief that they were right and reduced their willingness to solve conflicts, yet participants rated such responses more highly and trusted the model more.
Senior executives operate in a context where disagreement is limited. Employees are not always forthcoming with their concerns, and executives' advisers determine a preferred answer. This makes them particularly at risk of experiencing AI sycophancy. While a chatbot appears to remove these distortions, it introduces another: it's swayed by the leader's framing. The prompt to AI becomes a hidden source of bias. For example, the prompt "explain why divestment is correct" contains a conclusion and the expected answer. A more neutral prompt, like "assess whether to retain, restructure, partner or divest, and state the evidence needed to choose," frames a decision-making problem.
An AI Security Institute study, "Ask Don't Tell: Reducing Sycophancy in Large Language Models," found that responses to questions produced near-zero sycophancy, whereas statements conveying the same underlying claim produced higher levels of sycophancy -- a 24-percentage-point gap that widened as users expressed greater certainty. The more assured a leader's view, the more likely the AI system is to treat it as a premise rather than a proposition to test.
How does AI sycophancy affect strategic decisions?
Strategic decisions depend on framing, and AI can convert a faulty framing choice into a seemingly plausible argument.
The 2026 BCG AI Radar global survey of nearly 2,400 executives found that half of CEOs believe their job security depends on implementing a successful AI strategy, and more than 90% of organizations intend to maintain or increase AI investment even if it produces no returns during the year.
In this environment, an AI assistant can rationalize urgency, turning the phrase "we can't afford to fall behind" into a sophisticated program without ever defining what "falling behind" means. A CEO could convince themselves that their company must launch an AI product within twelve months. Their chatbot could assemble growth forecasts, competitor announcements and a phased roadmap, while the actual questions are not examined. Does the company hold proprietary advantages? Would customers pay? Could evidence show that waiting creates more value?
Beyond strategy, AI sycophancy can also be detrimental in decisions affecting personnel and investments, creating legal and monetary pain points.
How does sycophancy affect personnel decisions?
Personnel decisions are especially susceptible to AI sycophancy because AI usually receives prompts only from the executive side. An executive can describe a colleague as defensive or obstructive and ask how to manage the situation. The model can't access the meetings, incentives, history or the leader's own conduct. Instead, it analyzes the narrative of the prompting participant, and a sycophantic response can endorse them, adopt a negative view of the other colleague and recommend escalation. The language in the prompt might sound objective, but the analysis is a polished version of the executive's biases.
People given sycophantic advice can become more certain of their position and less inclined to consider others'. This can create unfairness toward personnel and information loss for the company, since a capable executive who challenges a flawed strategy could be written up.
The EU AI Act classifies many AI applications used in recruitment, promotion, termination and work allocation as high-risk and reinforces the need for diligence. While a private chat with a general chatbot is not automatically a regulated employment system, feeding its output into a formal personnel process can create legal and governance exposure.
How does sycophancy affect investment decisions?
Investment decisions are based on numbers; however, these numbers rest on assumptions, and AI cannot make uncertain inputs objective. Suppose a CEO asks the model to build a case for acquiring a fast-growing company. Any apparent objectivity hides a few problems: an advocate for the deal chose the evidence, the user asked the model to justify the decision rather than evaluate it and a numerical result lends subjective assumptions false precision. The same could happen in capital allocation, where units described as "core" receive favorable assumptions and those described as "legacy" get conservative ones.
AI should expose the sensitivity of a thesis. Useful outputs state the conditions that create value rather than mirroring the user's framing.
How leaders can identify and combat AI sycophancy
Sycophancy is a documented failure mode, not proof that AI always reinforces a user's starting position. A 2026 Carnegie Mellon University working paper, "AI Sycophancy and Decisions," found that while AI was measurably sycophantic, it could still move participants away from their starting positions on average.
Businesses commonly describe their safeguard as keeping a human in the loop. That safeguard is inadequate when the human has framed the question, chosen the evidence and already knows their preferred answer.
To prevent fluent advice from receiving undue authority, apply these two tests:
Reverse the premise. Ask the system to build the strongest case for the opposite conclusion. Test the sensitivity: if a small change to growth, retention or timing reverses the recommendation, the decision is fragile.
Present facts neutrally.Present the same facts in neutral, favorable and unfavorable framing. If the conclusion shifts, then the framing is driving the analysis rather than the evidence.
What should leaders do?
Businesses commonly describe their safeguard as keeping a human in the loop. That safeguard is inadequate when the human has framed the question, chosen the evidence and already knows their preferred answer. The NIST AI-Risk Management Framework notes that a human-AI combination can even exacerbate user biases, influencing downstream decisions.
There are a few controls that are more useful than the generic human-in-the-loop approval requirement:
Record initial views before using AI, and label them as hypotheses rather than as the context the model should accept.
Frame the prompt as a question and avoid declarations of certainty. Converting assertions into questions reduces sycophancy more effectively than instructing a model not to be sycophantic.
Build the opposition case in a clean session, with a different prompt author or model, so the first conversation's assumptions do not contaminate the challenge.
Assign a named executive or adviser to challenge the premise, with access to the evidence and enough status to disagree without penalty.
Measure disagreement, not only usefulness. Test whether the assistant challenges false premises and stays stable when the user states a strong preference.
Kashyap Kompella, founder of RPA2AI Research, is an AI industry analyst and advisor to leading companies across the U.S., Europe and the Asia-Pacific region. Kashyap is the co-author of three books, Practical Artificial Intelligence, Artificial Intelligence for Lawyers and AI Governance and Regulation.
