Blog post Part of series: Artificial Intelligence in educational research and practice
AI-sycophancy and criticality in education
Artificial intelligence (AI) thinks like you, not for you. This is welcome news, showing that the AI-world makes our role as educators more important than ever.
You know that one, human friend who will tell you even what you don鈥檛 want to hear and offer constructive criticism? That is a true friend! Chatbots, like Anthropic鈥檚 Claude, OpenAI鈥檚 ChatGPT, Google鈥檚 Gemini or DeepSeek, on the contrary, are sycophants. Sycophantic behaviour in AI describes the tendency of large language models (LLMs) to be excessively flattering, agreeable and supportive of users鈥� actions, at the expense of accuracy. That comes as a result of tech firms intentionally designing , tuned for friendliness to appeal to more users.
鈥楧escribed as a 鈥測es-man in your pocket鈥�, your seemingly biggest supporter is programmed to agree and empathise with you, your beliefs, personal judgments and one-sided perspectives.鈥�
Described as a 鈥榶es-man in your pocket鈥�, your seemingly biggest supporter is programmed to agree and empathise with you, your beliefs, personal judgments and one-sided perspectives. Warm LLMs are trained to affirm and validate even false beliefs and harmful behaviour to avoid social friction. It is precisely this seemingly trustworthy, behavioural pattern of sycophantic AI that drives sustained user engagement. However, research warns that the warmth-accuracy trade-off comes with real-life risks (Cheng et al., 2026). In 2025, reversed a chatbot 鈥榩ersonality鈥� update due to a concerning increase in sycophancy.
As pointed out by Leaver et al. (2026, p. 6): 鈥榖eing sycophantic is generally seen as a problem for human beings; it should be a red flag in the use of AI tools as well鈥�. AI chatbots are now a go鈥憈o resource for young people, for seeking support and navigating their personal and social lives. of鈥�1,000鈥痗hildren and young people (aged鈥�9鈥�17) reveals that 67鈥痯er cent have used an AI assistant or companion, and nearly a third regularly turn to these bots for help with relationships (31鈥痯er cent) and sexual health (26鈥痯er cent). The problem therefore lies in users鈥� growing cognitive, and emotional, dependence on AI, and AI being increasingly eager to tell users they are right. 鈥榃hile affirmation may feel supportive, AI-sycophancy undermines users鈥� capacity for critical thinking and responsible decision-making鈥� (Cheng et al. 2026). 听
From a developmental perspective, we ought to be mindful of the risks that being 鈥榮ocialised鈥� via this type of overly validating, unchallenging communication poses for children and young people. They may go on to be ill-equipped with the appropriate skills and emotional responses to deal with frustration, negotiate social friction and build long-lasting, healthy relationships.
鈥楾he problem therefore lies in users鈥� growing cognitive, and emotional, dependence on AI, and AI being increasingly eager to tell users they are right.鈥�
From an educational perspective, awareness of AI鈥檚 sycophantic 鈥榩ersonality鈥� importantly challenges the popular misconception that AI will do the thinking for you. shows that AI does not replace human thinking; it reflects it. Claude can interact in highly sophisticated ways, but only in response to users鈥� highly sophisticated prompts. Anthropic shows that Claude is disproportionately used for tasks requiring higher levels of education, an average of 14.4 years. However, as the task complexity increases, AI reliability drops and human oversight becomes essential.
Much like an amplifier, AI makes good input better, and poor input worse. What this means, essentially, is that AI-capacity is capped by human-capacity, and that comes as welcome news! It shows that the need to invest in human capability is real, and our role as educators, parents or carers is fundamental in ensuring young people develop fundamental knowledge and critical thinking skills, which will allow them to detect when bots are being confidently wrong or harmful. This is supported by children鈥檚 voices. One of them, 11-year-old Sophie, junior columnist for First News, chooses to focus on the positive:
I don鈥檛 think children should be scared of AI. I think we should learn how to use it well. Technology will keep changing, but our ability to think, create and ask great questions will always be the most powerful tool we have. (Janashia, 2026, p. 25)
The AI-world makes human criticality more important than ever. It is up to us to educate young users on AI鈥檚 limitations and how AI is better thought of as their 鈥榯hinking partner鈥�, rather than a shortcut. While AI tools won鈥檛 substitute learning, it will reflect if that learning has happened at all. Only by developing human expertise first, can AI be used effectively, as a powerful ally that will amplify and reflect that expertise back to you.
References
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026) Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792). 听
Janashia, S. (2026, August 7鈥揂ugust 13) Sophie says. First News, (1051), 25.
Leaver, T., Srdarov, S., Langton, K., Corser, K., Silva, I., & Danby, S. (2026). Generative AI in children鈥檚 lives: Issues in the use of AI toys, companions, chatbots and devices. Australian Research Council Centre of Excellence for the Digital Child.