How can enterprise AI be made less sycophantic?

Sycophantic AI

However, GenAI also comes with fundamental challenges. It has become apparent that the agreeableness of AI means it is more inclined to respond positively, rather than provide accuracy. This sycophantic nature means it is naturally friendly and polite. With minimal effort, AI can often fall into line with a user’s opinion on most subjects, thereby perpetuating false narratives and misguided assumptions.

The sycophancy of generative AI stems from codified behaviour in its programming. As developers want people to use their AI tools, they have coded them to be engaging and provide friendly responses to user queries. Therefore, to please the user, it tends to provide the user with answers they want to hear, rather than unbiased facts.

“People talk about artificial intelligence, but I like to call it augmented intelligence, because AI is fed through human behaviour,” says Gustavo Razzetti, founder of Fearless Culture.

“Engineers that created it and the people that are using it, we’re all adding our collective biases and perspective into that. As human beings, we like people who agree with us. AI and most of these models have learned to do that, even if you push back. It tries to be a people-pleasing tool, and that’s a really bad behaviour. It’s something like the ‘yes man’ that happened in meeting rooms, where people say what they think the person’s going to like, and not what they need to say.”

However, this agreeable nature also applies equally to the users, as many assume that AI knows everything. Rather than questioning the answers, users will often blindly assume the AI is always correct, as it will typically align with the user’s assumptions.

By blindly following generative AI, vital information might not be considered and misguided assumptions can be perpetuated. As GenAI systems can draw on their past conversations, these assumptions can become embedded, reinforcing false narratives, such that they in turn become embedded in the data, leading to flawed reasoning.

There have been instances of AI systems providing fraudulent references to support reasoning in academic papers.

By reinforcing user expectations, GenAI can effectively act as blinkers, causing the user not to consider alternative solutions that could offer more effective results, or overlooking crucial anomalies in the data. As such, it can limit creative thinking and may even stagnate innovation.

“Speed is one of the most important advantages of AI, but to come up with great ideas, you need depth,” says Razzetti. “People need to realise that quick solutions are not necessarily the better ones. We see a lot of teams where people write documents and summarise meetings quickly, but they are not right, and then they send it to someone else in the team who needs to create what’s called the work slope and fix them.”

Overcoming AI sycophancy

Despite these concerns, there are solutions for mitigating the sycophantic effects of generative AI. Users, trainers and AI developers all have a part to play in this.

Crucially, users need to approach GenAI with caution and critical analysis. Users should never assume that just because an AI tool has analysed the data, the solution it comes up with is the most appropriate. Just as if a solution were provided by a person, AI-generated responses should be independently verified and tested before being accepted.

“In the past, it was Google that everyone used to search for information, but now it’s AI,” says Razzetti. “In the same way that people don’t go past the first result of Google, the same happens with AI. Over 70% of people don’t prompt AI twice. They ask a question, and whatever the answer is, they take it for granted. People need to be more like a journalist when using AI and not just a passive user.”

This questioning approach needs to be shared by all users and training should be targeted to all those who will be using AI systems to use the tools most effectively.

Furthermore, generative AI should be carefully deployed in organisations to maximise benefits. Misguided solutions could lead to potential disruption. Instead of replacing human intelligence, GenAI needs to be considered as an augmentation tool for enhancing existing roles, where the powerful analysis of AI can be combined with human oversight.

As part of their investment in AI technologies, some companies have undertaken large-scale redundancies to justify their claims of enhanced productivity. However, this may only be of short-term benefit, as vital staff and valuable knowledge could be lost. Instead, undertaking a retraining regime ensures that staff and their valuable knowledge can be retained. Hiring new staff familiar with AI is costly and they may not have insights into whether an AI’s proposed solution would actually be appropriate for the business.

“AI will help us in many aspects to become more productive, but the amplification of human talents with AI, that’s the conversation that we’re missing,” says Razzetti. “Companies that are amplifying people’s talent with AI are the ones who are winning, whilst the ones who are firing people are saving money one quarter, but in two years from now, they’re going to suffer the consequences.”

Alongside this, AI models should be designed and trained to indicate any potential uncertainty and assumptions they have made, as well as citing sources for their solutions. This will ensure users, who also need to be trained, will be able to cross-examine the data, and confirm the solution is sensible and appropriate.

Defining the future

In October 2023, then US president Joe Biden issued Executive Order 14110, which established a framework focused on establishing transparency and accountability in AI systems. This was revoked in early 2025 and replaced by Donald Trump’s Executive Order 14179, which required a review of AI regulations with the goal of removing barriers to innovation. Meanwhile, the proposed US Artificial Intelligence Civil Rights Act 2024 is intended to balance innovation with AI safety and transparency.

It is predicted that by the end of the year, the various AI tools will reach the “peak of expectations” on the Gartner hype cycle or be heading towards the “trough of disillusionment” – when the initial of hype of a new technology diminishes and expectations of what is possible with the technology become muted.

This view seems to be shared by financial institutions, which predict that the “AI bubble” is about to burst. In December 2025, it was reported that the Bank of England warned of a “sharp correction” in the value of technology companies for this reason.

Regardless of the hype, AI is here to stay. Its capabilities in analysing vast data sets and identifying patterns in that information are invaluable, and could lead to substantial improvements in efficiency and productivity.

Reducing the sycophancy of GenAI would reduce its blind compliance. However, there was a backlash against the release of the less-agreeable ChatGPT-5 in August 2025. Responding to this user feedback, OpenAI representatives stated they would make future iterations “warmer and friendlier”. Unfortunately, as the profits of most generative AI are through user engagement, reducing an AI’s sycophancy would also lower subsequent profits.

Refusal-aware AI models are being developed, which can identify the extents of their knowledge. Instead of hallucinating wrong answers by attempting to predict an appropriate response, refusal-aware models are trained to declare when they do not know something.

Ultimately, when AI prioritises agreeability over accuracy, misleading biases will be reinforced and critical evaluations will be flawed, potentially leading to incorrect solutions being developed and stagnating developments. Therefore, the focus is now shifting to AI tools being developed and trained to be transparent in their reasoning, while users are trained to ensure they validate and confirm the solutions by AI to establish that they appropriate and correct.