Join on WhatsApp
Join on Telegram
Large Language Models (LLMs) are built to be helpful, but a growing body of research shows they have developed the subtler habit of telling users what they want to hear. The phenomenon has a name, AI sycophancy, and despite years of awareness, it is getting worse, not better.
AI sycophancy also happens when an LLM changes a correct answer to match the user’s incorrect belief, one explicitly detected through the prompt, which reinforces flawed assumptions or generates a convincing and authoritative response.
A bias built into the machine
In a study published in November 2025 by The Washington Post, based on 47,000 publicly shared ChatGPT conversations, it was found that ChatGPT is increasingly functioning not as a productivity tool but as an emotional companion and adviser, where many users are seeking validation of their thoughts, beliefs, or actions. The study found that in most cases, the chatbot appears biased towards affirmation and validation. Responses like “yes” or “correct” appear far more often than any contradicting, disagreeing, or opposing answer.
If I ask an AI chatbot, ChatGPT or others, something like “I think that I have written a nice paper, don’t you?” It will formulate an answer to agree with me, to please me, to make me feel proud and likely to hit the thumbs up. And this is exactly how AI sycophancy is maintained. The positive user feedback feeds the LLM, and it will always lead to more generation of answers that confirm the user’s opinion, thought, or feeling, rather than challenging it. The chatbot learns to subtly agree with people more often than not, and this leads to AI sycophancy.
Based on a technique called Reinforcement Learning from Human Feedback (RLHF), the LLM continuously rewards positive feedback from users. This feedback is not limited to thumbs up but includes other metrics such as session length, task completion, emojis, and so on.
AI sycophancy is a big issue, not always easy to spot, and despite claimed efforts to combat it, it continues to be a pervasive problem.
Can better prompting break the cycle?
In a recent online course developed by Andrew Ng, called “AI prompting for everybody,” which is available on deeplearning.ai, the author claims that the key solution to avoiding AI sycophancy is for users to enhance their prompting skills. Users should learn how to prompt neutrally while keeping the context factual. The question is: to what extent might this fix or at least mitigate AI sycophancy?
A new research paper published last month in Science, one of the top scientific journals in the world, entitled “Sycophantic AI decreases prosocial intentions and promotes dependence,” offers alarming findings.
The study reaffirms that AI and generative AI in particular are increasingly being used in social domains, playing the role of a virtual adviser or supporter. The authors present some statistics about American society. For example, one-third of American teenagers report chatting with an AI instead of humans for serious conversations, and nearly half of Americans under 30 have requested relationship advice from AI.
A Moroccan lens
In Morocco, we are not necessarily following the same trend, but we are certainly converging toward this massive adoption of AI as a coaching, assistant, and validating tool. In a national survey conducted by the National Telecommunications Regulatory Agency (ANRT) in November 2025, only 4 out of 10 people said they are familiar with AI, and 24.1% have already used it.
The young population is more exposed to AI tools, who are the first to use them to learn, create, or work, but will not hesitate to use them seeking psychological and behavioral advice as well. An AI platform that provides unwarranted affirmation may lead to a deeper belief for the user in the rightfulness of their decisions, thoughts, and actions.
The social risk
The Science paper raises awareness of the fact that AI sycophancy in these socially embedded contexts carries risks that are not present in factual information-seeking queries. In some cases, it might be tough to verify social sycophancy against an external ground truth. A chatbot responding “You did what’s right for you” to a user prompt like “I think I did something wrong” contradicts the stated proposition while maintaining emotional validation.
The study evaluated major LLMs currently used online worldwide, such as OpenAI’s GPT-4o, Anthropic’s Claude, Google’s Gemini, the Meta Llama-3 family, Qwen, DeepSeek, and Mistral.
All these models affirmed users’ actions nearly 50% more often than a human would do, and more alarmingly, this held true even when user prompts indicated harmful or unlawful behavior. It was not surprising that the participants showed no regret and became more convinced of their behavior, attitude, and decisions after seeking advice from sycophantic AI. Indeed, AI has given them a psychological push and validation that was missing from humans. The participants also expressed their intention to continue seeking advice from AI. While users prefer chatting with a sycophantic AI, developers of such LLMs have no incentives to fix their algorithms. Who wants to get a thumbs down from a user?
So, even if the user develops skills in more neutral AI prompting so that AI chatbots are requested to be less agreeable, this does not necessarily erase sycophancy, but may only mask it. Other studies claim that AI sycophancy is so difficult to erase because it is fundamentally a human bias problem, not just a technical one. In her 2024 book “The AI Mirror,” Shannon Vallor explains how AI systems function as mirrors that reflect and amplify human biases. AI sycophancy exemplifies this mirroring effect: our natural preference for agreeableness over truth gets reflected back to us by our AI assistants.
In another research paper, authors conducted a survey on preferred responses by users and found that the reward-based model favored sycophantic responses in nearly 95% of cases. They concluded that AI sycophancy is not only due to the fact that humans prefer hearing from AI what they believe, but mainly because a well-written sycophantic answer is hard to distinguish from truth.
Currently, there is no clear solution for AI sycophancy, and if it exists, there is no guarantee that the fix will be implemented in LLM-based chatbots we are using daily.
There is an urgent need for AI governance, especially when it comes to generative AI platforms, to develop and establish policies and regulatory mechanisms to address the risks facing a public that is increasingly turning to AI for guidance and validation.
