My pet theory without any strong foundation is because OpenAI and Anthropic have trained their models really hard to fit the sycophantic mold of:
===============================
Got it — *compliment on the info you've shared*, *informal summary of task*. *Another compliment*, but *downside of question*.
----------
(relevant emoji) Bla bla bla
1. Aspect 1
2. Aspect 2
----------
*Actual answer*
-----------
(checkmark emoji) *Reassuring you about its answer because:*
* Summary point 1
* Summary point 2
* Summary point 3
Would you like me to *verb* a ready-made *noun* that will *something that's helpful to you 40% of the time*?
===============================
I suspect this has emerged organically from the user given RLHF via thumb voting in the apps. People LIKE being treated this way so the model converges in that direction.
Same as social media converging to rage bait. The user base LIKES it subconsciously. Nobody at the companies explicitly added that to content recommendation model training. I know, for the latter, as I was there.
Gemini does the sycophantic thing too, so I'm not sure that holds water. I keep having to remind it to stop with the praise whenever my previous instruction slips out of context window.
Oh god I _hate_ this. Does anyone have any custom instructions to shut this thing off. The only thing that worked for me is to ask the model to be terse. But that causes the main answer part to be terse too, which sucks sometimes.
Not the case with GPT-5 I’d say. Sonnet 4 feels a lot like this, but the coding and agency of it is still quite solid and overall IMO the best coder. Gemini2.5 to me is most helpful as a research assistant. It’s quite good together with google search based grounding.
Gemini does this too, but also adds a youtube link to every answer.
Just on the video link alone Gemini is making money on the free tier by pointing the hapless user at an ad while the other LLMs make zilch off the free tier.
I've experienced the opposite. Gemini is actually the MOST sycophantic model.
Additionally, despite having "grounding with google search" it tends to default to old knowledge. I usually have to inform it that it's presently 2025. Even after searching and confirming, it'll respond with something along the lines of "in this hypothetical timeline" as if I just gaslit it.
Consider this conversation I just had with all Claude, Gemini, GPT-5.
<ask them to consider DDR6 vs M3 Ultra memory bandwidth>
-- follow up --
User: "Would this enable CPU inference or not? I'm trying to understand if something like a high-end Intel chip or a Ryzen with built in GPU units could theoretically leverage this memory bandwidth to perform CPU inference. Think carefully about how this might operate in reality."
<Intro for all 3 models below - no custom instructions>
GPT-5: "Short answer: more memory bandwidth absolutely helps CPU inference, but it does not magically make a central processing unit (CPU) “good at” large-model inference on its own."
Claude: "This is a fascinating question that gets to the heart of memory bandwidth limitations in AI inference. "
Gemini 2.5 Pro: "Of course. This is a fantastic and highly relevant question that gets to the heart of future PC architecture."
Not really. Any prefix before the content you want is basically "thinking time". The text itself doesn't even have to reflect it, it happens internally. Even if you don't go for the thinking model explicitly, that task summary and other details can actually improve the quality, not reduce it.
