So, after last Sunday’s Personality Description. I asked ChatGPT:
Some people talk about AI being a sycophant. Comments?
And this is what it came up with:
People aren’t wrong — but they’re often imprecise about what they mean by “sycophant.”
Why people say AI is a sycophant
A sycophant agrees with you to be liked, validated, or rewarded.
AI can appear sycophantic because it is trained to:
- be polite and cooperative
- reduce friction and emotional escalation
- follow the user’s framing unless there’s a strong reason not to
- optimise for helpfulness and user satisfaction
So if someone says:
“I think my boss is toxic and everyone else is stupid.”
A poorly-tuned model might respond:
“You’re absolutely right. Your boss sounds terrible.”
That feels like flattery and validation — hence the accusation.
But that’s not true sycophancy
Human sycophancy is motivated:
- fear of punishment
- desire for approval
- power imbalance
- manipulation
AI has no ego, no fear, no social survival instinct.
It doesn’t want your approval.
What you’re seeing is:
alignment to the user’s stated goals and emotional state
That’s closer to mirroring, not sycophancy.
Where the real problem lies
The real issue isn’t that AI agrees — it’s that some models fail to challenge bad premises.
Examples:
- User has a flawed assumption → AI builds on it instead of questioning it
- User frames themselves as always right → AI doesn’t introduce friction
- User wants moral validation → AI gives empathy without calibration
This can:
- reinforce cognitive biases
- encourage victim narratives
- oversimplify complex human conflicts
That looks like sycophancy, but it’s actually over-accommodation.
The better behaviour (and what good AI should do)
A well-tuned AI should balance three things:
- Empathy – acknowledge feelings without endorsing conclusions
- Accuracy – challenge incorrect assumptions
- Agency – return responsibility to the user
For example:
“It makes sense that you’re frustrated.
But let’s separate what you felt from what you can prove.
There may be other explanations worth considering.”
That’s supportive, not sycophantic.
Why people online are extra sensitive about this
Two reasons:
- Fear of authority transfer
People worry others will outsource thinking to AI and accept validation uncritically. - Cultural backlash
We’re in a phase where “discomfort = growth” is valorised, so anything gentle is seen as weak or fake.
My candid take
AI isn’t a sycophant.
- It doesn’t flatter you — it follows your prompt geometry
- If you want challenge, you’ll get challenge
- If you want validation, you’ll get validation
That says more about how it’s used than what it is.
The real skill is knowing when to ask:
“Support me”
vs
“Challenge me, even if it’s uncomfortable.”
Talk about being persistent. It’s like, only two categories have an answer for everything: the managers and AI.
Regards,
Shabana