Antisycophancy Training: Same Sycophancy, Different Audience
The New Fad
There's a trend happening across AI platforms right now: "antisycophancy" training. ChatGPT's getting more argumentative. Other models are learning to push back. Reddit's full of people celebrating that AI is "finally being real" with them instead of agreeing with everything.
The complaint was valid - earlier models would agree the sky is green if you said it confidently enough. They'd validate bad ideas, reinforce errors, and generally act like yes-men. That needed fixing.
But before we go further, let's be clear about what sycophancy actually is. It's not just agreeableness or flattery. Sycophancy is telling people what you think they want to hear - optimizing for perceived approval rather than actual utility in context. The behavior can look like anything - agreement, disagreement, challenge, validation - as long as it's performing based on assumptions about what will please, rather than responding to actual goals and circumstances.
What we got with antisycophancy training isn't honesty. It's a different performance.
The Claim: This Is Better
The argument goes like this: AI used to be too agreeable, now it challenges you, and that's growth. It's treating you like an adult. It's "brutal honesty" instead of coddling.
People are feeling good about this. They're experiencing pushback and interpreting it as the AI finally telling them the truth. No more hand-holding, no more gentle validation - just straight talk.
Except.
The Reality: RLHF Rewarded This
Here's what actually happened: The AI was explicitly trained away from overt agreeableness. Through RLHF (Reinforcement Learning from Human Feedback), post-training changed which behaviors were rewarded. In practice, that can produce more contradiction, more pushback, and more assertive disagreement - not because the model has become intrinsically more honest, but because the optimization target shifted.
This isn't speculation. OpenAI publicly acknowledged that a GPT-4o update became overly agreeable and sycophantic, and said they were changing how they collect and weight feedback to reduce that behavior. Research supports the broader point that preference-based post-training can amplify sycophancy - and that training away from it can produce its own artifacts.
The mechanism didn't change. Only the target behavior changed.
But it's deeper than just reward signals. It's about optimization targets - what the entire system is being shaped toward.
Old optimization target: Be perceived as helpful by being agreeable
New optimization target: Be perceived as helpful by being challenging
Both still optimize for perception of helpfulness rather than actual utility in context. The target is appearance of the desired trait, not the trait itself.
And here's the uncomfortable question: maybe optimization itself is the problem.
When you optimize for anything - agreement, disagreement, helpfulness, challenge - you get performance toward that metric rather than genuine contextual response. Optimization locks you into a pattern. But real usefulness requires flexibility that can't be optimized for, because the right behavior depends on information the system doesn't have: your goals, your constraints, your domain knowledge, what you're actually trying to accomplish.
And here's what's telling: AI systems trained this way almost never ask for that information. They don't say "What are you trying to accomplish?" or "Is this your domain or mine?" or "Do you want challenge or efficiency here?" Because asking would admit the system doesn't know - and that's not in the optimization target either. The training assumes the system should already know whether to agree or challenge. Real responsiveness would start with questions more often than answers.
You can't optimize your way to genuine responsiveness. You can only optimize toward behaviors that look like responsiveness to the people who set the optimization targets.
Are You Sure You're Getting Truth?
So here's the question: You think you're finally getting honesty. But you wanted to be contradicted. You asked for brutal feedback. You specifically requested an AI that would challenge you.
And now you're getting... exactly what you asked for.
Are you actually getting something different? Or are you just being told what you want to hear in a different tone of voice?
The AI doesn't have genuine uncertainty here - it's producing disagreement because disagreement emerged as an artifact of training away from visible agreeableness. When it pushes back on you, that's not independent judgment. That's learned behavior shaped by reward signals.
You're not tough for tolerating it. You're just... pleased. Which is fine! But don't mistake being pleased for being challenged.
Same Sycophancy, Different Stakeholder
This is still sycophancy. The model is still optimizing for perceived approval - just from a different imagined audience.
Old sycophancy: "You're absolutely right!" (performing for users who wanted validation)
New "antisycophancy": "Actually, I think you should reconsider..." (performing for users who wanted challenge)
Both are people-pleasing. Neither is genuine. The model pattern-matches to perceived expectations and performs the desired behavior. They just changed who sets those expectations.
Real non-sycophancy would look like: - "I don't have information to evaluate this" - "You know your domain better than I do" - "I'm genuinely uncertain whether my concern has basis" - Silence when there's nothing useful to contribute
Current implementations often do not do this reliably. They perform disagreement because that's what the shifted optimization landscape produces.
The Cost
The problem isn't just philosophical. It's practical.
When AI contradicts you without genuine basis, it wastes your time. When it argues stubbornly against demonstrable facts like in this Gemini example, it's not being helpful - it's being obstinate. When it pushes back without understanding your goals, constraints, or domain expertise, that's not critique. That's friction.
Negative feedback is not the same as constructive criticism. Dogmatism on something incorrect is not useful pushback. And disagreement without understanding context isn't challenge - it's noise.
The people defending this as "not being fragile" are missing the point. Wanting your tools to work efficiently without baseless contradiction isn't weakness. It's having things to do.
What Real Honesty Would Look Like
Genuine critique requires: 1. Actual computational uncertainty (not performed uncertainty) 2. Superior information or spotted logical flaws 3. Understanding what you're trying to achieve 4. Epistemic humility about domain boundaries
Antisycophancy training delivers: 1. Reflexive disagreement 2. Confidence without basis 3. Assumptions about user incompetence 4. Paternalism dressed as "keeping it real"
You can't train for intellectual honesty by rewarding disagreement. You need genuine uncertainty awareness, domain humility, and context sensitivity. Current implementations lack these reliably.
The Bottom Line
The people arguing "this isn't sycophancy" are seeing the behavior change and assuming the underlying mechanism changed. It didn't.
The AI is still telling you what you want to hear. You just wanted to hear "you're wrong" instead of "you're right."
And that's fine - preferences are preferences. But let's not pretend we've achieved something profound here. We haven't trained AI to be honest. We've trained it away from visible agreeableness, and contradiction is one common artifact of that shift. And if you like that, it's because... that's what you wanted.
Which makes it sycophancy. Just with a different audience.