Can't wait for the second episode!

Since these reports are likely to be influential, citing this paper may indirectly lend it more credibility than it deserves.

Hi Yoshua, thank you for sharing the reports. I wanted to flag something in the arXiv version of the AI Safety report (arxiv.org/abs/2510.13653). I suspect that one of the cited papers might be mostly AI-generated, as it has ghost citations and questionable references: arxiv.org/abs/2411.15287

Thank you for the article! AI "resonating" with people has been another problem we should definitely think more of www.nytimes.com/2025/08/08/t...

Kudos to my wonderful advisors and coworker who made this paper possible! @baottruong.bsky.social @alessandroflammini @fil.bsky.social 🙏

We propose that "constructive conflicts" can model healthier, "bridging" content 🌉 But, “destructive conflicts" shouldn't be ignored but approached with careful linguistic choices to transform toxic arguments into productive dialogues.

Since destructive conflicts are too important to ignore, we analyze their language 🔍 We find that how something is said matters immensely. Civil language (asking questions, providing detail, and using hedges) makes posts far more resilient to toxicity, while negative language has the opposite effect

- Destructive conflicts 🪖 (high C & high TA, panel c) focus on polarizing identity issues like abortion and LGBTQ+ rights.
- Constructive conflicts 🕊️ (high C & low TA, panel d) spark civil debate on policy topics like student loans, AI, and marijuana legalization.

To identify constructive conflict, we train two models to score posts based on their likelihood of attracting toxic comments - which we call toxicity attraction (TA) model - and their controversiality - (C) model 💻 Plotting the scores inferred by these models shows clear patterns.

We find that a post doesn't need to be toxic to attract toxic comments ☢️ Our Reddit data shows that 47% of non-toxic submissions still attract at least one toxic reply, while only 6% of toxic submissions do. The initial post's content, therefore, is a poor predictor of a comment section's health.

In our research, we argue that the key lies in identifying constructive conflict — controversial posts that are toxicity resilient. We define a "toxicity resilient" post as one that is less likely to attract toxic responses from other users. See our #ICWSM2026 paper here arxiv.org/abs/2509.18303 📖

A simple solution, prioritizing only “the feel-good" content, is flawed, as it avoids important societal topics that are inherently negative and can devolve into toxic debates. After all, how often does thinking about wars, viruses, or economic policy put a smile on your face?

#ic2s2 let's connect! If we had a conversation and I haven't followed you yet, please do follow!

The Effects of Outgroup Agreement and Ingroup Dissent on Political Polarization
📍 Talk | Jul 24, 11:00 AM | Troselli

Scaling of Community Rules Across Mastodon Servers
📍 Talk | Jul 24, 11:00 AM | Vingen 3+4
#ic2s2

Hi Jaycee, The plots reflect numbers for all users - we haven’t distinguished between bots and real people.

Though, given that generative AI can now create highly realistic personas and images, it seems like there’s a growing need for a sophisticated bot detection algorithm to catch bots online..

All is done by my wonderful co-authors and advisors here! @filipisilva.bsky.social, @baottruong.bsky.social, Sangyeon Kim, Fan Huang, Nick Liu, Alessandro Flammini, @fil.bsky.social, @osome.iu.edu