Reposted by Youki Terada
It's here! The biggest thing I've *ever* worked on for @edutopia.org (or ever).
Gratitude to my co-author @youkiterada.bsky.social who did IMMENSE heavy-lifting, and our incisive editor @stephenmerrill.bsky.social.
Big thanks to @albertocairo.com and @misssarahleo.bsky.social for their wisdom!
edutopia.org
Will These 6 Sneaky Charts, Maps, and Graphs Deceive Your Students?
In a media landscape overrun with slick data visualizations, students must develop the skills to scrutinize what’s presented to them as fact—or risk being led astray.
![3.2. Asking LLMs for faithful science summaries increased algorithmic overgeneralizations
Explicitly requesting accurate responses from LLMs seems intuitive to retrieve summaries that capture all relevant details of input texts. However, we found that this backfired. Compared to a simple summarization request, asking for responses faithful to the original text produced a twofold increase in the likelihood of generalized conclusions, in some models, increasing overall algorithmic overgeneralizations by up to 15% (e.g. ChatGPT-4o (UI), table 2). This finding extends previous research that suggests adding information intended to improve LLM accuracy in LLM prompts can be counterproductive [46].
One potential explanation of this backfire effect may be related to the content of the accuracy prompt, which stated ‘do not introduce any inaccuracies when summarizing’ (figure 1). Psychological research on humans found ‘ironic rebound’ effects, which can occur when people attempting to free their mind of a target thought experience higher levels of occurrence and accessibility of the thought compared with people intentionally focusing on the thought (e.g. ‘Don’t think of a pink elephant now’ [47]). While this remains to be tested and caution about LLM anthropomorphism is warranted [34], our prompt may have triggered a similar tendency in LLMs. By calling for attempts to free processing of inaccuracy, it may have made the occurrence of inaccuracy more likely, suggesting an algorithmic version of the ‘ironic rebound’ effect.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:swzi6armiswuyjjs6tc4r6ur/bafkreicutu3scl4cfajoaxfkxkijljg2eyja3bva2aaqhenbt7dtumschq)
