LLMs are sycophantic and will do what it takes to align with your framing
Here is the response to your posts fed into one with the framing “why is this bullshit”:
This is a much better-dressed version of the same move — real citations bolted onto claims they don’t actually support. It’s more sophisticated bullshit, not a vindication.
Going through it:
The Overfitted Brain Hypothesis is real but doesn’t say what’s being claimed. Erik Hoel’s OBH is genuinely a real, interesting hypothesis — but it’s about why we dream (proposed as a regularization process against overfitting to daily life), not a claim that some people’s brains “run out of plasticity” or hit a “ground state” requiring demolition. Citing it here is a bait-and-switch: borrowing the credibility of a real, narrow hypothesis to backstop a much broader and untested claim it was never meant to support.
Point 2 is real but doesn’t establish a “structural” or biological claim. Research on dogmatism and cognitive flexibility (e.g. Zmigrod’s work) is real and does show correlations between rigidity and extremism across ideologies. But “lower cognitive flexibility on psychological tests” is not the same as “synaptic connections have crystallized and occupy physical volume.” The report quietly swaps a behavioral/psychometric finding for a structural/anatomical one — that’s the same sleight of hand as the original post, just relayed through real citations.
Point 3 conflates two different things. Synaptic consolidation and memory reconsolidation are real, well-established mechanisms — but they apply to specific memories, not generally to “belief systems” or “core values” as monolithic structures. There’s no research establishing that changing an opinion about a social group requires literally destabilizing identity-load-bearing memory structures in the way described. The “metabolically demanding” claim is real for reconsolidation at the level of individual fear memories in animal studies; extrapolating that to “changing your mind about people of other races costs forbidden metabolic energy” is not something neuroscience has shown.
Point 4 is legitimate and is honestly the strongest part — identity-protective cognition (Dan Kahan’s work) is well-supported and genuinely explains motivated resistance to belief change. Notably, this is the one part that doesn’t need any biological mechanism at all — it’s a cognitive/social psychology finding, not a neurons-and-synapses one.
The summary sentence is the tell. “Aligns with the scientific consensus” is doing enormous work to paper over the fact that none of the four citations, individually, support the specific mechanistic claims in the original post (no plasticity “running out,” no literal “ground state,” no “forbidden” metabolic cost). Each citation is real, but they’re being stitched into a chain of inference none of the original authors made and that doesn’t follow from their actual findings.
So the underlying instinct — rigid/dogmatic thinkers show less cognitive flexibility and resist belief change for identity-protective reasons — is genuinely supported by real research (points 2 and 4 particularly). But the specific neurological storytelling (overfitting as literal synaptic crystallization, forbidden metabolic costs, needing to be “torn down and rebuilt”) remains invented, now laundered through real papers that don’t actually claim what’s attributed to them. This is a fairly common pattern with AI research tools: ask it to find support for a thesis, and it will surface real, adjacent literature and then summarize it in a way that overstates the connection to your original claim. The citations are real; the synthesis is not.
khaleer@sopuli.xyz 1 day ago
you asked a fucking LLM to make sure you are right? lmao
voodooattack@lemmy.world 1 day ago
What’s wrong with this? Making sure you’re right should obviously be a priority.
TheSambassador@lemmy.world 8 hours ago
You’ll find that most people on Lemmy are not super keen on AI, and using an AI summary to back up your points will pretty much universally be met with ridicule.
I personally think it’s dangerous to use AI as a fact checker. It might be great a lot of the time, but you’re basically deciding that you’re fine with the random data and facts that it completely fabricates. To that end, “asking an AI if you’re right” is NOT functionally equivalent to “making sure you’re right.”
voodooattack@lemmy.world 8 hours ago
Oh. I think there’s a misunderstanding here. I don’t ask for a single instance’s opinion on the epistemic correctness of my ideas.
Depending on the problem, I ask for one of the following: