yikes

"America, in other words, isn’t illiterate. It’s postliterate." Tough but interesting long read on the American erosion of literacy www.theatlantic.com/magazine/202...

@katestarbird.bsky.social is my hero. @ic2s2.bsky.social

Great paper from Duncan Watts: LLM-based content reranking improves personalization & aligns video recommendations more closely to users’ prior viewing. Very interesting tho unclear if there is significant deployment of LLM-based recommenders systems at present? infodem.upenn.edu/research-com...

Late to work because I wrote an annoyed email to the NPR Politics Podcast hosts, but seriously the coverage was ridiculous

An awkward ED thing is you write more staged quotes for PR purposes but I AM ACTUALLY honored that @penn-mediated.bsky.social has joined @thegni.bsky.social And we're grateful to Jason Pielemeier, Montserrat Legorreta, and all the GNI selection folks for their support!

Writing about automated content moderation - it's amazing how many papers were focused on toxicity, with the comparison point being the Perspective API The sunsetting of the API means a loss for reproducibility And, this will probably be the case for many papers based on commercial LLMs

Trying to get people to read summaries of our new research and seeing I'm competing with @tressiemcphd.bsky.social saying she will deadlift her enemies and tbqh I suspect I'm not on her level

So this March poll from @prl.bsky.social finds lower usage among Americans for politics and election information, relative to @dataforprogress.org poll of U.S. _voters_ 46% use AI for news infrequently tho @seanjwestwood.bsky.social @ylelkes.bsky.social americaspoliticalpulse.com/report/trust...

The we're-paying-too-much-attention-to-Twitter-and-not-enough-to-YouTube problem, quantified New pre-print from @dfreelon.bsky.social @davidlazer.bsky.social @pranavgoel.bsky.social @meredithpruden.bsky.social et al osf.io/preprints/so...

Firefox security team wrote up a very interesting post saying they have radically accelerating security vulnerability fixes with the use of Anthropic's Mythos model. They also get into the process of how they integrated LLMs into a pipeline for hardening software hacks.mozilla.org/2026/05/behi...

Full table. Interesting gender (women more skeptical) and educational (college-educated less skeptical) numbers

Yikes, this @dataforprogress.org from February reports that 39% of U.S. voters said they are very likely or likely to use AI chatbots to learn about candidates or elections. Big age disparity, little partisan diff. Can't say this fills me with hope. www.dataforprogress.org/blog/2026/2/...

Ok one more - maybe my favorite single takeaway - from this paper on Multilingual fact-checking using LLMs The authors discover that model accuracy is negatively correlated with the amount of internet content available in a given language. ...bad sign for the internet...

Also notably that when claims are phrased as beliefs or opinions, it becomes much harder for LLMs to identify them as false - really interesting failure state.

Several papers find the same Dunning-Kruger-like effect Larger LLMs perform factchecking better, with less confidence, while smaller models perform worse, but more confidently.

Using RAG on well-curated and high-quality context significantly improves performance. Simply prompting models, by comparison, goes pretty badly. Interestingly, web search doesn't help very much on its own, sometimes adding low quality info (mixed/conflicting results across different studies)

In updating our LLM and Civic Discourse dashboard, I read 20+ papers on LLMs for fact checking (use the misinfo + fact-checking filter). Lots of interesting takeaways, some promising, some discouraging, but sufficient progress to make me think that LLMs in the future will be useful in fact checking

"We normally change 90K but we'll charge you 60K and can't tell you why" is hardly befitting of the most important institutions for economic mobility in the U.S. Anyway I wrote about this back in the day @techpolicypress.bsky.social and have recommendations: www.techpolicy.press/enrollment-a...

It's worth engaging w/ the first contention of the Yale Committee Report: universities aggressively adopting personalized pricing has significantly reduced trust in higher ed: president.yale.edu/sites/defaul... This number is striking: half of Americans think college costs are the same for everyone

An interesting note that I'd missed about Australia's Social Media Minimum Age law is that they are running a multi-year study following 4,000 families to evaluate the impacts of the age requirement on mental health, digital habits, family relationships, etc www.esafety.gov.au/research/soc...

Hope you enjoyed! Chile is a spectacular country. And if it was too much to read, no worries, I'm as apathetic as this llama (or yama, as they say, and which is obviously the more fun pronunciation). 9/9

8/9

7/9 (seriously so much better than rain)

6/9

5/9

4/9

3/9 (with for sure my favorite photo of the day)

2/9

Ok, it's my birthday and I don't that much going on (teaching a yoga class and a small dinner at Laser Wolf) But I thought it'd be fun to share the adventure I went on a year ago today, hiking to the Parinacota Volcano in Lauca National Park in remote northern Chile 1/9