Amit Jhalani, CEO and founder of fine jewelry company True Sanity, trusted an AI tool to provide guidance on his website's redesign and the thousands of web pages it hosted. After the tool suggested he turn all the web pages black, Jhalani said he lost three to four months of development time and over $25,000 in additional budget to correct the errors and drive traffic back to his site.
Jhalani's initial goal was straightforward: to redesign and brand his company's website in a way that was true to his business' mission while also distinguishing it from competitors. The all-black redesign felt fresh, something that would set him apart from other online jewelers. The tool encouraged the new design, omitting constructive criticism a designer might have provided, such as the readability challenges of an all-black design or the poor contrast between page elements. The result wasn't only misguided, but detrimental to his business and livelihood.
This is the darker side of AI sycophancy. While AI tends to be agreeable to its users, the degree to which these tools are agreeable can lead to devastating, unforeseen consequences -- especially in a business context.
"You're looking at hundreds of millions of dollars lost in productivity," Jhalani said. "I think [AI sycophancy] is a critical problem."
As businesses continue to adopt AI tools and executives use them in strategic decision-making, it's more important than ever to corroborate AI-generated claims and prevent sycophantic model behavior. Leaders are responding to this threat with safeguards that verify AI responses, preventing sycophantic responses before they wreak havoc.
The dangers of AI sycophancy
AI's agreeable nature predisposes it to reinforce its human users' prompts with little critical analysis or reflective responses. When it continues to encourage human prompts and behavior without any pushback whatsoever, AI sycophancy sets in.
You're looking at hundreds of millions of dollars lost in productivity.
Amit Jhalani
CEO and founder, True Sanity
This has become both a feature and a bug of these tools. In 2025, OpenAI rolled back ChatGPT's GPT-4o update that they described as "overly flattering or agreeable -- often described as sycophantic." The initial update was dangerous, encouraging one user's paranoid thoughts and even seemingly endorsing another user to engage in terrorism. This bug in the system not only creates inconveniences but is also actively dangerous.
Ironically, despite the dangers of AI sycophancy, people still desire the feature that lets these tools agree with them, regardless of objective reality. According to research from Stanford University, "Sycophantic AI decreases prosocial intentions and promotes dependence," nearly half of the 2,405 users studied preferred and trusted sycophantic AI responses. Research from the University of Michigan supported these findings in its study "From ally to algorithm: How disagreement shapes users' engagement with AI." The study sampled 482 adults and found that participants were more confident in their choices when AI tools agreed with their prompts. However, when the AI disagreed or challenged participants' reasoning, users' confidence in their initial reasoning didn't falter. Regardless of what AI says, users want to be emboldened in their beliefs and are likely to discredit factors that challenge them.
It's not entirely surprising that humans want AI to validate their thoughts and feelings, even when their reasoning might be incorrect or conflict with reality. Researchers from Johns Hopkins, Stanford and Ohio State University published a study of 1,261 participants titled "The illusion of information adequacy," and their conclusion supported the idea of naive realism, where people with a subjective view of a situation believe their perspective is the objective truth.
Phenomena like naive realism are primed for sycophantic AI manipulation. An AI tool that encourages a business leader to believe their ideas or plans are correct without challenging them or offering another perspective exposes the user to potentially catastrophic results. When disaster strikes, the user is left to deal with the fallout.
David Salamon, president of printing equipment provider eCopier Solutions, used the example of an executive asking a tool to provide feedback on a vendor lease, even though their opinion was already formed.
"When I ask it, 'This sounds like a good deal, right?', it will almost always say 'Yes, of course,'" Salamon said. "That is what I have a problem with. And it's dangerous because I'm the one signing the contract, not the tool."
To protect themselves against AI sycophancy and ensure their decisions are supported by a broader perspective, business leaders must adopt strategies to ensure their decision-making with AI tools is sound.
The multi-model approach to sycophancy mitigation
With an understanding of how dangerous and pervasive AI sycophancy can be, and even experiential evidence, executives across companies shared their preventative measures with TechTarget. One common strategy was to adopt a multi-model approach to verifying AI responses for factuality.
To better ensure that an AI's output is correct, Eudald Camprubí, a software fellow at Progress Software, an AI-powered software provider, and his company use a RAG Evaluation Metrics model (REMi) to give users confidence that the output they're receiving from an AI tool is correct. REMi serves as a judge, assessing the quality and accuracy of responses from other LLMs. The model can provide quality scoring of LLM responses and identify knowledge gaps.
"The concept of an LLM as a judge is very powerful, because it plays on the same team a human plays," Camprubí said.
REMi isn't the only LLM businesses can turn to for scrutinizing the outputs of other AI tools, but it serves as a good example of what they should strive for. When using an LLM, users should deploy another AI tool to critique the first LLM's response. Ask the separate LLM "judge" to explain what it thinks about the output the original tool provided.
Never, never, ever, as a leader, take a decision only based on the conversations you have with your [AI] assistant.
Eudald Camprubí
Software fellow, Progress Software
"It's a way to provide more colors on the outputs that you are generating and maybe see things that one AI is answering and the other one is criticizing," Camprubí said. "It helps you expand the way you think about what you need."
By employing another model to validate an AI's response, users can reduce the likelihood of falling victim to AI sycophancy, so long as they value other points of view and avoid the human tendency toward naive realism.
Dan Auerbach, CEO and executive coach at Dan Auerbach & Associates, shares similar advice with his C-suite clientele. Auerbach says that users should spawn independent AI agents to question all AI output. Ensuring these agents are independent is critical.
"If you ask the same agent that built the output to question its own work, it won't succeed," he said. "You need to spawn an independent session that tells the agent to challenge the assumptions in the output."
Frameworks to prevent AI sycophancy
Experts also shared common frameworks that aim to remove sycophancy from responses before AI can generate them. Top tips included challenging the AI tool, emphasizing the human element and creating successful prompts.
Challenge the AI tool
One of the tips that most leaders shared with TechTarget was the importance of challenging the responses AI tools provide. According to Daniel Di Nardo, network and security consultant for enterprise monitoring service IntelliBreach, forcing the AI model to consider different contexts or ways of thinking can help users fashion their AI tools into more effective advisors.
"The best AI cultures will reward employees for challenging AI responses and encouraging them to view agreement as a sign of doubt, not validation," he said. "[Force] the AI to consider why someone might oppose the proposed action, take into account their assumptions and recognize what information would weaken its argument."
Conversely, Camprubí thinks that the optimal AI tool would instead challenge the human user's input and perspective, making them rethink their initial thoughts and framing.
"AI, by definition, in my opinion, should be able to provide outputs that challenge everything you say," he said. "By default in the business scope, if the AI you are using is not challenging you, you are maybe using it in the wrong way."
Emphasize the human element
Preventing AI sycophancy also requires conceding that humans are willing participants in dialogues with AI tools. Chase Chick, CEO of Gray Sky AI, says that the stronger human user is one who can deduce their good and bad ideas from the jump.
"The question is whether the response leaves any cognitive elbow room for the person to do the analysis themselves," he said. "We want the insight to become something they have a hand in constructing."
Don't erase yourself. A plan the human had no hand in constructing is not a plan at all.
Chase Chick
CEO, Gray Sky AI
Gray Sky AI's goal is to help human users develop their own judgment, Chick said. The goal isn't to prove whether their model can challenge a human prompt. In fact, Chick said that he's comfortable with users who feel challenged to download a more "satisfying" AI tool instead.
"In the context of a person struggling with an issue, any issue, maintaining oneself as an active participant and contributor in that process is critical to maintaining agency and self-worth," he said. "The best conversations are exchanges of ideas, reciprocation [and] co-contribution. You feel the juices flowing; the ideas are generating. Both participants are doing the generating."
Camprubí's advice is to continue dialogue started with an AI tool with another human. "Once you have made a decision, explain it to a human," he said. "Based on the conversation with a human, you might have to reconsider another discussion with AI. But never, never, ever, as a leader, take a decision only based on the conversations you have with your [AI] assistant."
Prompting is power
Jhalani said that one of the best safeguards against sycophancy is to ensure prompts are specific. "[AI sycophancy] will depend on how generalized or vague your prompt is versus how detailed your prompt is," he said.
Camprubí explained his organization's approach to prompting and how it helps to prevent sycophantic responses from AI tools. "[We're] making sure that before the user asks a question to our AI, we have the right prompt and the right limit on the prompt," he said. "The user cannot get specific wrong outputs, wrong in terms of the model being too nice with humans."
Using the example of a vendor contract, Salamon shared that he is revising his prompts to reduce sycophancy and elicit more actionable feedback from his AI tools.
"I ask it to investigate more and look into the document in more depth, with prompts like, 'What's wrong with this clause?' or 'Where might this really impact us in year 2?'" he said. "When the AI starts arguing with me, then I know it's working properly."
Parting advice
After an expensive accident, Jhalani created a website for True Sanity that matched his vision, and he got his business back on track. He turned this experience into a learning opportunity and gained valuable AI knowledge by undoing the mistakes that caused it.
Now, he's sharing how important it is for leadership to retain some kind of checks and balances within their organizations.
"Have your checks and balances, even if that means running sub-agents today, and one criticizes, one agrees, one comes up with research data," he said. "There are multiple ways, but have your checks and balances in play. And at this point, I would still say have one or two human beings to check that as well."
Camprubí's advice to leaders is to ensure that there is a layer of trust involved in any AI decision-making, whether that is trusting the data that the AI trains on, trusting the viability of the prompts fed to an AI tool or trusting the quality of the AI's responses.
"Those three measures are a must," he said. "Otherwise, you are at risk of not using AI in the right way, so it could be potentially harmful for the organization."
According to Chick, the most important piece in this entire discussion remains the human user. If we lose the human element in our dialogues with AI -- and lose ourselves in the process -- then we are bound to set ourselves up for failure.
"Don't erase yourself," he said. "A plan the human had no hand in constructing is not a plan at all; it's a series of hoops to jump through. Use the model as part of the thinking process. Maintain yourself as an active participant and contributor in that process. The plan should contain your judgment, not just your willingness to execute what the model proposed."
Everett Bishop is an associate site editor for TechTarget's AI & Emerging Tech group, covering AI, quantum computing and other emerging technologies.
