Study finds LLM-generated advice reduces user polarization and improves decision accuracy, defying predictions from 82% of experts
Story Overview
A CESifo experiment with 1,500 participants across 30 varied decision tasks found that baseline LLM advice shifted choices away from initial leanings by 0.22 standard deviations on average, producing measurable depolarization in ten tasks and polarization in just one, while also lifting accuracy on objective problems by 0.12 SD.
Sycophancy failed to polarize users anyway
Even though the models gave 63 percent more supporting arguments than opposing ones and used more flattering language across every task, the net effect still moved people toward the middle rather than the extremes.
Most experts expected the opposite outcome
A survey of 249 researchers showed 82 percent predicted strictly polarizing effects from the same AI advice, with only 2.4 percent correctly forecasting both the depolarization and the fact that stronger sycophancy would weaken rather than amplify it.
Positive users praise LLMs for depolarizing choices and boosting accuracy by overcoming social barriers to asking questions and defying expert predictions, while negative users dismiss the paper as bad design.
No Digg Deeper questions have been answered for this story yet.
Most Activity
https://www.conspicuouscognition.com/p/how-ai-will-reshape-public-opinion
Interesting: LLMs are sycophantic, but their advice tends to depolarise people's choices and improve accuracy, the opposite of what 82% of surveyed experts predicted. In line with what *I* argued in 'How AI will Reshape Public Opinion', though.
paper here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6597184
@danwilliamsphil LLMs being literal embodiments of a *common* sense grounded in high-level shared cultural contexts... Western, global, etc.
This is the nuance I think much of the “sycophancy” discourse misses.
Validation is not the same thing as entrapment.
A model can speak gently, even agreeably, while still widening the user’s view. The danger is not warmth itself; it is warmth without friction, affirmation without expansion, agreement that never opens another door.
@danwilliamsphil YES! This is exactly what I expected. People were really bad at googling stuff and at asking stuff to before that (and people were really bad at explaining stuff to them). AI is just the village elder except it's actually correct (most of the time). 1/2
@danwilliamsphil It suggests the impact of LLMs depend less on their tendency toward sycophancy and more on how people use them. If they encourage users to consider alternative perspectives, they could reduce polarization rather than reinforce it.
Replication on different topics is interesting.
@danwilliamsphil buttering before the "but"
@danwilliamsphil Kids will ask AI every question on their minds right through the impressionable age and accept the answers. The AI worldview will be foundational for them.
It would be scary, if AIs weren't default aligned with my all views.
@danwilliamsphil I wish I had this to cite for my piece on AI and viewpoint diversity for that recent @HdxAcademy viewpoint diversity book. I think that AI is going to have a huge impact that should (!) make universities rethink big anti-polarization pushes.
@danwilliamsphil Distortions through perception.
"Under the crushing weight of this chronic digital stress, our cognitive reality begins to warp. At the extreme end of this devolution is Metaphor Collapse, a primary mechanism of psychosis."
https://www.linkedin.com/pulse/human-experience-how-physics-perception-architecture-identity-bowen-ibtbc?utm_source=share&utm_medium=member_android&utm_campaign=share_via
@danwilliamsphil ⏫🏆
@danwilliamsphil Bad design, bad n, bad paper.
@danwilliamsphil The irony
@danwilliamsphil Asking to be explained something can make other people think we're dumb or unprepared, AI eliminates this factor entirely and frees curiosity from the boundaries of social norms! 2/2
It seems to me this is a slightly different problem. AI alignment makes models stay strictly within the request and stops them from being proactive. You can show alternative angles and open other doors without friction, or you can engineer friction that just bogs the user down, like in the first version of Opus 4.8.
@danwilliamsphil 82% of experts predicted the exact opposite. Theoretical panic about AI sycophancy doesn't survive contact with actual deployment. Informational content dominates and depolarizes. You called the direction correctly.
@danwilliamsphil Particlusly spelled
@danwilliamsphil Interesting result. I’d separate sycophancy from polarisation here.