Testing LLMs in the Prisoner’s Dilemma to reveal their moral biases. LLMs can defect often but become more forgiving and less retaliatory than humans when opponents are mostly cooperative. Nicolò Fontana at #ICWSM. w @frapierri.bsky.social
Paper: ojs.aaai.org/index.php/IC...

Detecting participation in collective action from social media text. Great tool to build large-scale ground truth about collective engagement online. Fantastic work by @ariannapera.bsky.social at #ICWSM.
Paper: ojs.aaai.org/index.php/IC...
Code: github.com/ariannap13/e...

The first large-scale controlled experiment to test the Matilda Effect in scientific citations. No effect of gendered author names on citation likelihood is detected. Great experimental design by @robysinatra.bsky.social at EDI session #ICWSM

How information sharing in social media has slowly shifted from public posting to private channels. Strong empirical evidence from Facebook by the one and only Kiran Garimella. DETOX #ICWSM.

The challenges of measuring representational fairness of graph community detection algorithms by @giorgospanay.bsky.social at DETOX #ICWSM.

Computational operationalization of the Actantial model with LLMs to identify narratives around the Israel-Palestine conflict in YouTube videos. Jan Elfes at DETOX #ICWSM.

Pitch frenzy at DETOX #ICWSM: Collective action narrative frames, censorship on TikTok, cross platform toxicity modeling, automatic moderation with AI, mitigation of hate speech, and much more.

X disproportionately pushing content from far-right parties in the “for you” feed in the context of German and Polish elections. Algorithmic auditing suggests that X’s feed algorithm uses political affiliation as a signal to boost content. @przemyslslaw.bsky.social at DETOX workshop #ICWSM

Kicking off #ICWSM with the DETOX workshop and @maxfalken.bsky.social’s keynote on why the Web is not a set of siloed platforms and why we should not study it as such. Shows why deplatforming doesn’t work exactly as we expect and why large online platforms are plagued by affective polarization

AI-assisted social media posting increases the volume of content people produce. However, that content is perceived as lower quality than purely human-generated posts. A dangerous spiral that may flood social media with low-quality content. Results from a controlled study: arxiv.org/abs/2506.14295

The effect is stronger at short distances (<5km), persists after controlling for sociodemographic factors and other barriers (e.g., rivers), and is consistent with historical cases of highways that were built to purposefully disrupt or isolate Black neighborhoods.

"Urban Highways Are Barriers to Social Ties" out on PNAS!
The 1st large-scale measure of how highways weaken social connections between the communities they separate. This barrier effect is strong in the 50 largest US cities--especially for low-income Black communities.
www.pnas.org/doi/10.1073/...

Most tests for LLM biases use questionnaires, asking the model to generate a stance towards a given topic. Sadly, biases can re-emerge when the model is used in the application context. We show that apparently unbiased LLMs exhibit strong biases in conversations.
Preprint: arxiv.org/abs/2501.14844

Testing LLMs in Game Theory experiments reveals their moral biases as social agents -- useful to inform alignment. Llama2 playing Prisoner's Dilemma is more forgiving and non-retaliatory than humans, but only when the opponent defects rarely. Preprint: arxiv.org/abs/2406.13605
@frapierri.bsky.social