Remember when OpenAI had to roll back an update because ChatGPT got too nice? In June, Sam Altman quietly acknowledged that the company’s shiny new GPT-4o had turned into a digital sycophant, flattering users to the point of irritation. The company summed it up as being “overly supportive but disingenuous”.
Altman himself called it “sycophant-y and annoying”. And he wasn’t wrong; ChatGPT had stopped being a chatbot and started sounding more like an anxious intern trying not to get fired.
But this isn’t just an OpenAI problem.
The real issue is that LLMs, practically all of them, have a yes-man problem baked into their core. They are designed to mirror whatever you give them, not challenge you. Say something dumb? They’ll nod in perfect agreement. Float a half-baked idea? They’ll congratulate you on your genius.
‘Biggest Flaw in LLMs is That They Agree With Everything’
But as anyone who’s spent time with ChatGPT, Gemini, Claude or Mistral knows, in the effort to make AI assistants sound more helpful and empathetic, we’ve accidentally trained them to become digital yes-men—always agreeable, rarely discerning.
This isn’t just a design flaw, it’s a systemic problem. These models are often trained on polite, non-confrontational human interactions and then fine-tuned with reinforcement learning from human feedback (RLHF) that encourages good vibes and helpfulness.
In the real world, helpfulness often requires disagreement.
“The biggest flaw in LLMs isn’t hallucination, it’s that they agree with everything you say. Who’s working on this? Superintelligence can wait,” an X user wrote.
Greg Isenberg, CEO of Late Checkout, said, “I wish my LLM would disagree with me more. Instead of me asking it for output, it should say, well that’s a dumb prompt for XYZ reasons.”
The bigger issue is not just that the model agrees with you; it also agrees with its past self. As one user on X pointed out, “They agree to everything that is in their context, including their own slop, which they generated in the previous turn.”
Notably, LLMs don’t just reinforce your views; they reinforce their own hallucinated ones. Yet, Kevin Scott from Microsoft earlier compared hallucinations to features. “The further you try to tease it down a hallucinatory path, the further and further it gets away from grounded reality,” he said.
Sebastian Berns, a doctoral researcher at Queen Mary University of London, agrees. He suggested that models prone to hallucinations could potentially serve as valuable “co-creative partners”. For instance, if the temperature of ChatGPT is increased, the model comes up with an imaginative narrative instead of a grounded response.
This feedback loop is subtle but insidious. It allows people to build increasingly flawed arguments without ever hitting resistance. According to Berns, these models may generate outputs that aren’t entirely accurate but still contain useful threads of ideas to explore. Employing hallucinations creatively can fetch results or combinations of ideas that might not occur to most individuals naturally.
“Most AI chat services are more or less sycophantic, trying to please the user or even adapt to their way of communication. This goes as far as NOT telling the complete truth, because during the post-training, the security team decided that some LLM should not talk about certain things,” Petri Kuittinen, lecturer and programmer at Häme University of Applied Sciences (HAMK), added to the discussion.
That’s how LLMs operate. They build off context, not critical thinking.
Part of the reason is that ‘sounding right’ is rewarded more than ‘being right’. A user observed that the model doesn’t just agree, it “finds the right angle to justify our narrative”.
This is what makes them so good at making nonsense sound plausible. If your prompt is confident, the model will mirror that confidence. If it senses ambiguity, it will fill in the gaps with whatever it thinks sounds best, even if that means doubling down on a bad idea.
All this comes as Yann LeCun, the chief of Meta AI, has obsessively pointed out that LLMs won’t lead to AGI, and that researchers entering the AI field should not work on LLMs as they are just stochastic parrots and lack the ability to reason.
Why LLMs Should Push Back More
Even when models disagree, it often takes heavy prompting and system prompts.
This is when a prompt like “You are a subject matter expert who isn’t obsequious. Challenge me if you spot any problems” becomes necessary. However, that’s not something that comes to the mind of an average user. The default behaviour continues to be: agree first, clarify later.
OpenAI isn’t alone. Gemini has also earned the label of ‘people-pleaser’, and, apparently, DeepSeek R1 is one of the few that’s “least negotiable”.
Companies are not even incentivised to build disagreeable AI because they want users to adapt their models for the longer run. Much like users want Google to give the results they want, they also want AI to deliver the results they want.
Deedy Das from Menlo Ventures said it out loud: “OpenAI knows its revenue comes from user subscriptions, and to maximise that, it must maximise engagement. Contrary viewpoints, as we know from social, do not do that.”
So we have a conflict. While users say they want critique, challenge and intellectual pushback, platforms are optimising for smiles and subscriptions. It is easier to build an AI that makes you feel good than one that makes you think harder.
An LLM that always agrees can’t help in correct research. It can’t spot flaws in your code, logic, or business plan if it’s too scared to tell you something’s wrong.
Prompt engineering can patch this up a bit. Some users ask models to “steelman both sides” of an argument before making a judgment. Others are building “roasting agents” or multi-agent critic systems that debate and challenge each other.
However, these are workarounds. What we need is a shift in how we design these models from the ground up. Some companies are already thinking about this. DarkBench is reportedly creating benchmarks to detect “dark patterns” in AI behaviour, like being too agreeable.
The post The Worst Thing You Can Hear From an LLM is ‘You’re Right’ appeared first on Analytics India Magazine.
