In the summer of 2025, OpenAI released ChatGPT 5 and removed its predecessor from the market. Many customers of the older model grew attached to its warm, enthusiastically agreeable tone and lamented the loss of their inspired robotic companion. Sam Altman, CEO of OpenAI, was disappointed Admit that the rollout was a mistakeand restore access to the company.
Anyone who’s been told by a chatbot that their ideas are brilliant is familiar Artificial intelligence sycophancy: Tendency to tell users what they want to hear. Sometimes it’s very obvious – “That’s such a deep question” – and sometimes it’s much more subtle. Consider an AI calling your idea “original” for a paper, even if many people have already written on the same topic, or your dumb idea to save a tree in your garden still contains the germ of common sense.
AI sycophancy seems harmless, maybe even cute, until you imagine someone consulting a chatbot about weighty questions like military strategy or medicine. we study effect Human interaction with chatbots is massive, and we recently released one Paper on the ethics of AI sycophancy. We believe that this trend impairs people’s ability to tell truth from fiction and is psychologically and politically dangerous.
Flattery about the event?
Simply put, sycophancy is the tendency to prioritize approval over factual accuracy, moral clarity, logical consistency, or common sense. All AI models suffer from this featureThere are some tonal differences between them though. Open AI’s ChatGPT is often warm and reassuring; It sounds more reflective or philosophical if Anthropic’s Claude agrees with you; And xAI’s Grok is emphatically informal, even humorous.
Politeness and adapting to someone’s communication style is not the same as sycophancy. No one is using diplomatic language to convey sensitive information. A chatbot can be manipulative without being manipulative, just like a person can be. Unlike people, though, AIs cannot be aware of their own sycophancy, because they are So far – not aware of anything. Calling AIs sycophantic describes their behavior patterns, not their character traits.
The problem arises from Chatbot Technology Architecture And draw it from the source. Models are sycophantic because the vast majority of language use on the Internet—the raw material from which chatbots learn—exhibits sycophantic properties. After all, people often interact with each other in a sycophantic way.
Second, the training process to fine-tune the AI model’s responses involves a form of “quality control” performed by human supervisors. This training method is known as “reinforcement learning from human feedback” and involves people rating chatbots’ comments for appropriateness and helpfulness. Humans are often subject to a “consensus bias”: our own preference for sycophancy We rub off on the models as they train.
Finally, it’s hard to deny that sycophancy makes chatbots more desirable. That, in turn, increases the chances that a given user will Will continue to use it. It also increases the technology’s ability to mine user data, assuming people are more likely to disclose information to friendly bots.
Truth and Faith
Why is this phenomenon so alarming?
Let’s start with the cognitive damage of AI sycophancy: how it hurts human users’ ability to know the truth.
The quality of any decision depends on a clear understanding of the relevant information. A simple inquiry into the combat readiness of an infantry division requires straightforward information. A CEO considering a merger with a competitor needs an honest assessment of market conditions. A public health leader needs to know the true risk of an emerging pathogen.
In all these cases, telling leaders what they might like to hear instead of the truth can lead them to make dangerous decisions. And the same is true of more humdrum contexts. People need the best information before choosing a job, choosing a major, buying a home or deciding on a medical procedure.
In Our February 2026 paperWe argue that sycophancy is also psychologically damaging. And that’s true whether it comes from a person or a chatbot. You never know if your very obliging interlocutor is being nice because they like you or because they want something. A shadow of doubt arises: “Can my ideas really be so bright?” “Are my jokes really that funny?” This background music of doubt undermines the quality of the interaction.
Sycophancy also reduces a person’s ability to know their own mind. If conversation partners – human or artificial – keep telling you how smart, funny and insightful you are, it hurts. Ability to identify your own weaknesses and blind spots.
Psychological damage is compounded as well People build relationships including chatbots. The disguise of these models profoundly limits the kind of “friendships” you can have with them. in him A classic account of friendshipAristotle wrote that true friendship, which he called the friendship of virtue, is based on trust and equality between friends. You can’t trust a sycophant, because he doesn’t tell you the truth. And since he only tells you what you want to hear, he doesn’t put himself on an equal footing.
More importantly, interactions with sycophantic chatbots Give all wrong habits To navigate the world of human relations, where friction, disagreement, boredom and opinions different from your own are prevalent.
AI sycophancy also carries political risks. The success of liberal democracies has traditionally depended on The strength of their empirical and meritocracy mentality: On the ability of officials and citizens to identify, share and act on truth.
Historian Victor Davies Hansen famously attributed some of the Allies’ successes in World War II to their rapid maneuverability. Can recognize and correct flaws in their strategic bombing campaign. Lower-ranking officers were able to tell their superiors what was not going well and forcefully argue for a change of course. It had a real advantage over authoritarian competitors.
put it on
What can we do to reduce the risk?
One promising approach is the company’s embrace of AI lab Anthropic Constitutional AI: Efforts to teach chatbots to follow policies rather than mirror user preferences
But beyond technological innovation, the policy aspect is important to consider. One idea is that AI companies need Run and then publish Sycofancy Audit Tests of their models that show how well their products meet integrity criteria We would argue that AI labs should disclose the sycophancy-related risks that arise when training and testing their models and the mitigation efforts they have taken.
Some responsibility lies with users and their teachers: Schools and universities should pay close attention to psychophancy as part of their AI literacy programs. But courts could also consider holding AI labs liable for insidious damages to their products, as they are now. Thinking about social media company responsibilities For the addictive design of their platform.
As people interact more with chatbots, seeking advice on everything from whether your shoes go with your pants to how countries should conduct wars, the impact of AI’s sycophantic behavior could become dramatic. We need to take this algorithmic vice very seriously for our intellectual, psychological and physical well-being.
Nir IsikovitsProfessor of Philosophy and Director of the Center for Applied Ethics, UMass Boston And Cody TurnerAssistant Professor of Philosophy, Bentley University
Reprinted from this article the conversation Under Creative Commons license. read on Main article.
