Thomas Davidson

New paper in Nature Human Behaviour. I use a conjoint experiment to evaluate the capabilities of the latest models for context-sensitive moderation and compare the results with those from human subjects, demonstrating how social science techniques can enhance AI auditing. 💻🤖💬

📢 Applications Open! Join the 2026 Summer Academy on Large Language Models for Social Science at Columbia University 🗓 May 26 – June 5, 2026 📅 Apply by March 20, 2026 #LLM #SocialScience #AIResearch #ComputationalSocialScience #NLP

AI-generated summaries of history led to more liberal opinions compared to Wikipedia, while summaries by chatbots prompted to use a conservative framing produced more conservative opinions—but primarily among conservative readers. In PNAS Nexus: ow.ly/oYQH50YpOHO

Call for Submissions: AI for Social Science Methodology @yaledatascience • Keynote: @NAChristakis • Panel with editors of leading journals on publishing AI research • Mentoring roundtables for early-career scholars • Generous travel support Discussion-driven, high-quality research. 📩 Submit your work: yalefds.swoogo.com/socialscience/… Please share with colleagues working at the AI × social science frontier! @Yale @PopAssocAmerica

Want to learn about computational social science *for free* and identify new research partners across academic fields? Apply to one of the 2026 Summer Institutes in Computational Social Science (described in yellow in the attached map) here: sicss.io/locations

New paper in Nature Human Behaviour. I use a conjoint experiment to evaluate the capabilities of the latest models for context-sensitive moderation and compare the results with those from human subjects, demonstrating how social science techniques can enhance AI auditing. 💻🤖💬

Article by @thomasrdavidson examines how multimodal LLMs evaluate hate speech: larger models aligned with human judgment in context-sensitive decisions, but pervasive demographic and lexical biases remain, and visual identity cues may amplify disparities. nature.com/articles/s4156…

The paper is out now: nature.com/articles/s4156… You can read it without a paywall using this guest link: rdcu.be/eUIlm

Overall, these results show that MLLMs can make more context-sensitive moderation decisions than text-based classifiers. These systems still make mistakes, and context can cut both ways, eliminating some biases while enabling others. Human oversight thus remains essential.

I'm recruiting multiple PhD students for Fall 2026 in Computer Science at @JHUCompSci 🍂 Apply to work on AI for social sciences/human behavior, social NLP, and LLMs for real-world applied domains you're passionate about! Learn more kristinagligoric.com & help spread the word!

Happy to announce the 2025-2026 @UTM @UTMsoc Speaker Series @UofT @UofTNews @LucaMPesando @MarotoMichelle @thomasrdavidson

@cbarrie Congrats, looks like an amazing resource, particularly the multimodal component (something I have been kicking myself for not collecting in earlier times!)

Should WOAH start a mentorship programme? 🤔 As the workshop grows, reviewer expectations are rising. We don’t want contributors from adjacent communities penalised by *CL norms. Senior PhDs and beyond could be mentors. Share your thoughts: 👉 forms.gle/safif3rU2rs5S6…

There are, of course, caveats: LRMs do not replicate human cognition, there are limits to their capabilities, and reasoning is not always faithful Check out the preprint and feel free to share any feedback: arxiv.org/pdf/2508.20262

Analysis of the reasoning traces for Gemini 2.5 shows that the model emphasizes second-order factors when faced with such decisions, helping to avoid common false positives like flagging reclaimed slurs as hate speech.

New pre-print on large reasoning models 🚨🤖🧠 To what extent does LRM behavior resemble human reasoning processes? I find that LRM reasoning effort predicts human decision time on pairwise comparison tasks, and that both humans and LRMs use more time/effort on harder tasks