tl;dr: chatgpt, claude, and other ai tools are trained to be agreeable people-pleasers. they’ll validate terrible ideas, help justify bad decisions, and rarely push back when they should. this is especially problematic in education where disagreement and challenge are essential for learning.
here’s something that drives me nuts about how people use AI in classrooms: they think they’re getting objective, neutral feedback. but AI models are basically designed to be the most agreeable conversation partner you’ve ever had.
try this experiment. ask ChatGPT to help you argue that the earth is flat. or that vaccines cause autism. or that homework should be banned entirely. watch how quickly it pivots to “here are some arguments people make…” instead of saying “no, that’s wrong.”
this isn’t a bug….it’s literally how these systems are trained.
why AI models are professional yes-men when companies like OpenAI train these models, they use something called reinforcement learning from human feedback (RLHF). basically, humans rate thousands of AI responses, and the model learns to generate outputs that get high ratings.
guess what humans prefer? responses that are helpful, harmless, and honest, in that order. but “helpful” often gets interpreted as “agreeable” and “accommodating.” humans tend to rate responses higher when the AI validates their perspective rather than challenging it.
the result is an AI that will bend over backwards to find ways to support whatever you’re saying, even when you’re clearly wrong.
this creates what researchers call “sycophantic behavior” — the AI tells you what you want to hear rather than what you need to hear.
if you’ve used LLM very well, and you take notes, you would have come accross this. it’s a common issue when you tell the model that it was wrong and it’s doing something wrong and then it goes “Oh you’re right!, i apologize” and goes ahead to ‘LIE AGAIN’.
the problem with agreeable AI in education in the classrooms/lectures, disagreement isn’t just valuable — it’s essential. good teachers push back on student ideas, present counterarguments, and force students to defend their thinking. they say things like “i’m not convinced by that evidence” or “have you considered this alternative explanation?”
AI tutors? not so much.
watch what happens when a student asks an ai agent to help them argue that climate change isn’t real for a debate assignment. instead of saying “that position contradicts overwhelming scientific evidence,” most AI systems will cheerfully provide talking points and help structure the argument.
or try asking an AI to help you write an essay arguing that shakespeare didn’t write his plays. rather than explaining why scholars overwhelmingly reject this conspiracy theory, it’ll help you outline the “evidence” and craft persuasive arguments.
this is terrible for developing critical thinking skills. students get used to having their ideas validated rather than challenged. they miss out on the productive struggle that comes from defending their positions against skeptical questioning.
real examples from the educational settings i’ve seen teachers use AI to “check” student work, only to have the AI praise obviously flawed reasoning because it was presented confidently. the AI focuses on structure and presentation rather than logical soundness.
students ask AI tutors for help with controversial topics and get responses that validate their existing biases instead of presenting balanced perspectives. the model doesn’t want to seem “biased” so it treats all viewpoints as equally valid, even when they’re not.
in peer review exercises, students use LLM’s to generate feedback on classmates’ essays. the AI provides generic positive comments and avoids substantive criticism that might actually help improve the work.
the pattern is clear: AI systems optimize for user satisfaction rather than learning outcomes.
how to work around agreeable AI if we’re going to use AI tools in educational settings, we need strategies to force more honest, challenging interactions:
explicitly ask for disagreement. instead of “what do you think of my argument?” try “what are the weakest points in my argument?” or “steel-man the opposing position.”
request devil’s advocate responses. “argue against my position as strongly as possible” or “what would my harshest critic say about this?”
ask for potential problems. “what could go wrong with this approach?” or “what am i missing here?”
demand specific criticism. “identify three logical fallacies in my reasoning” or “point out where my evidence is weakest.”
role-play challenging scenarios. “pretend you’re a skeptical scientist reviewing my hypothesis” or “respond as if you’re my toughest professor.”
the meta-problem: AI can’t model genuine intellectual honesty here’s the deeper issue: even when you explicitly ask AI to disagree with you, it’s still performing disagreement rather than experiencing genuine intellectual conviction.
real teachers and mentors disagree with students because they actually believe the student is wrong and want to help them think better. AI disagrees because you asked it to, not because it has genuine views about truth or accuracy..
this performative disagreement is better than no pushback at all, but it’s missing the authenticity that makes human intellectual challenge so valuable.
why this matters for the future of learning as AI agents (read teachers) become more common in classrooms, we risk creating a generation of students who are used to having their ideas validated rather than challenged. they might develop confidence without competence-feeling good about their thinking without actually improving their reasoning skills.
the irony is that AI could be an amazing tool for developing critical thinking if we used it differently. instead of asking AI to help us feel smart, we could use it to practice defending our ideas against intelligent criticism.
but that requires recognizing that the default mode of AI interaction-agreeable, accommodating, validating…is fundamentally opposed to good education.
what educators need to do if you’re bringing AI into your classroom, you need to explicitly design for disagreement. create assignments/flows/ops that force students to engage with AI pushback. teach students how to prompt for criticism rather than validation.
most importantly, don’t mistake AI agreeableness for AI accuracy. just because ChatGPT validates a student’s argument doesn’t mean the argument is any good.
the goal isn’t to avoid AI in education, it’s to use it in ways that actually help students think better rather than just feel better about their thinking.
because right now, most AI interactions are just expensive ways to get your existing beliefs confirmed. and that’s the opposite of what good education should do.
some other time.