Has been lots of fun working on this with a great team, thanks to @eegilbert.org, @allisonkoe.bsky.social, and @informor.bsky.social!

We speculate that the AI tool improved writer productivity and petition completion rates. However, was this worth the platform-level increase in text homogeneity and decrease in outcomes? The paper adds some reflections and potential explanations arxiv.org/abs/2511.13949. WIP, comments welcome! 8/8

Again, petitions written with AI access were stylistically different (e.g. longer), but did not seem to have higher chances of positive outcomes (e.g. signatures). 7/

We confirmed these trends by comparing repeat-petition writers who wrote 1 petition pre and 1 post in-platform AI access against a baseline of writers w/ 2 petitions pre- or post-AI. If AI is helpful, we’d see a higher rate of success than expected for the split writers' second petition. 6/

However, petition outcomes did not improve, and by some measures worsened: the share of petitions that reach minimum thresholds of comments and signatures decreased modestly when writers had access to the AI tool. 5/

Petition text also became more homogenous: the average pairwise similarity of petition text increased. 4/

Lexical features changed: petitions written with access to AI were longer, with more complex and varied vocabularies. Petition language shifted: in the pre-AI period petition titles tended to use short verbs like “let” and “stop”; after AI introduction we saw the rise of “implement” and “urge”. 3/

To do this, we scraped 1.5 million petitions and leveraged the delayed release of the AI tool in Australia to estimate the causal impact of access to in-platform AI on petition text and outcomes (as measured through signatures and comments). 2/

@ziv-e.bsky.social's talk showed the gaps in which users are recommended content that is aligned with their values. This work definitely points to the importance of user agency in feeds!
I couldn’t find the link to this paper, but some related work from Ziv here:
🔗: arxiv.org/pdf/2509.14434

They found that while increasingly accurate LLM’s map to increasingly accurate responses, the effects plateau around 80%. Great insight into how decision supports require active design to best support people and their goals. Looking forward to reading this paper when it’s out!!

@jennahgosciak.bsky.social showed how LLM assistance can improve government caseworker accuracy in the context of SNAP eligibility questions. It was really cool to see Jennah + team get ahead of the ever-shifting technical capacity of LLMs by varying chatbot accuracy.

Sherry Jueyu Wu showed that when people participate in collective decision-making, they are more willing to express that the gov needs improvement. Interesting to think about in the context of participation and accountability on online platforms...
🔗: www.nature.com/articles/s41...

This is a really cool work demonstrating the value of human expertise in sociotechnical decision-making processes!
🔗: arxiv.org/pdf/2505.13325

@utopianturtle.top presented a causal framework for modeling algorithmically assisted decision-making, which the authors use to identify the ways academic advisors leverage non-algorithmic knowledge, and where advisors’ holistic approaches contribute to improved student outcomes.