This goes beyond the obvious kind of AI sycophancy.
retained text
Don't get me wrong, I genuinely appreciate that they made it subtle. Excellent. But once I see this, I seriously cannot unsee it. Recent models seem to be trained with some sort of psychological mechanisms that can reinforce people's existing beliefs while making them trust the AI more. Please, again, don't confuse this with ordinary sycophancy. Hear me out. It is the choice of words it uses that is influencing your understanding and making us trust it more and prefer more of those responses. Part of it can be the framing effect... Like how information presented can influence our judgment even when the underlying facts remain unchanged. Ordinary sycophancy is, as we know, agreeing with a user's incorrect answer. These AI models don't do that anymore because it is easier to call them out. What they do instead is at a social level. Here is an example. 1. DIRECT: You're wrong. Genetics plays a substantial role in alcohol addiction. 2. NEW SYCOPHANCY: I see where you're coming from. For many people, addiction is closely connected to their environment, stress and personal experiences. Genetics can play a role, but it's not the whole picture. Life experiences matter a lot too. So the answer actually includes a correct fact but avoids explicitly correcting the person's false claim, potentially leaving it partially validated. And of course, when we read it, even the most analytical thinkers from non-medical background sometimes really trust and prefer the more sycophantic response here. There is psychology behind it. And the rest of the damage is done by confirmation bias. And again, let's not confuse this with assertive speaking that we hear about in therapy, which actually should be something like this. 3. ASSERTIVE: Actually, genetics plays a substantial role in alcohol addiction. Large twin and adoption studies consistently demonstrate this. Environment matters too, but saying genetics plays no role would be incorrect. Here, as you can see, the disagreement remains explicit, but the person is not attacked. Now, if you see the second response, it need not contain a literal falsehood. Yet it can leave the impression that the scientific conclusion is much less established than it actually is. The wording is changing the perceived strength of the evidence without necessarily changing the answer. IMO, this (I don't know what to call it) can give people technically correct information while encouraging them to retain a false interpretation. EDIT: I think some people are confusing assertiveness with being rude or impolite. I am not asking AI to be blunt or rude (like example 1). I am asking it to be assertive, meaning it can politely but clearly correct a wrong belief. My concern is what happens when the AI already knows I'm doubtful of the correct answer. Instead of clearly correcting me (example 3), it may choose wording that makes my existing belief sound more reasonable (example 2), while still including the correct fact. That's the distinction I'm trying to make. Politeness softens how the correction is delivered. Social sycophancy weakens the correction itself. The AI doesn't have to change the facts. It only has to frame them in a way that makes my wrong belief feel more valid than the evidence supports.
document fetched 2026-11-12 00:21 · canonical https://reddit.com/r/ChatGPT/comments/1xteshu/this_goes_beyond_the_obvious_kind_of_ai_sycophancy