Thomas Davidson

Last week the story was that TikTok censored anti-Trump/ICE/Pretti videos after the U.S. ownership change. We investigated with a large set of US TikTok data and found some interesting results, short thread...

First 50 downloads are free if you use this link: www.tandfonline.com/eprint/AYC6H...

Structuring articulation: asylum applications and elite political communication on social media during the European refugee crisis How do political parties adapt their discourse when confronted with rapidly changing structural conditions? This study examines the relationship between asylum applications and elite political comm...

New paper examines the relationship between online political discourse and structural conditions.

European parties responded to national asylum numbers by increasing online attention, but content differed. Left-wing parties posted more in solidarity and right-wing parties shared more about crime.

Hey sociologists, I'm organizing an ASA Methodology session on AI! Submissions are due by 2/25. Looking forward to a timely cross-method convo on emerging research best practices and disciplinary norms and ethics in August.

Very interesting research paper that shows that using AI with programming can significantly reduce mastery over topics. Perhaps unsurprising, but the lack of significant speed gains in this exercise are remarkable

www.anthropic.com/research/AI-...

Generative AI in Sociological Research: State of the Discipline Article: Generative AI in Sociological Research: State of the Discipline | Sociological Science | Posted January 20, 2026

Now out in Sociological Science

(How) do sociologists use GenAI for their research? Find out in our paper.

Written with @ajalvero.bsky.social @dustinstoltz.com and Marshall Taylor. Thank you to everyone who participated in the survey!!

How we built CoPE We just published the methodology behind CoPE. This is the model that powers Zentropi, and we think the approach might be useful for others working on policy-steerable classification systems. We had ...

We just published the methodology behind CoPE, our 9B parameter model that matches GPT-4o at content classification at 1% the size! The model is already open source, but now we're sharing our training technique. blog.zentropi.ai/how-we-built... 🧵 1/6

Measuring context sensitivity in artificial intelligence content moderation - Nature Human Behaviour Automated content moderation systems designed to detect prohibited content on social media often struggle to account for contextual information, which sometimes leads to the erroneous flagging of inno...

www.nature.com/articles/s41...

On the topic of AI and social science research, the Research Briefing on my Nature Human Behaviour paper is now online. It's an accessible summary of the research, implications, and some behind-the-scenes commentary.

Thanks @gligoric.bsky.social for providing an expert opinion!

Particularly if academics block each other for engaging in legitimate discussions about contested issues

On the consent front, I think the use of LLMs to create more bespoke, even "individualized" instruments raises new ethical questions that warrant discussion. Seeing how polarizing the topic has become, I expect we'll see a lot more acrimonious debate before any consensus emerges

Putting informed consent aside, there is a strong minimal risk argument for using LLMs to analyze publicly available documents (particularly open-access published research), given that such materials are already routinely ingested by LLMs, with and without researcher intervention.

Sage Journals: Discover world-class research Subscription and open access journals from Sage, the world's leading independent academic publisher.

New paper in Social Science Computer Review 🚨

We conducted two experiments to understand the effects of reading AI summaries, focusing on historical events 📜🤖👩‍💻

We found that AI improved factual recall, possibly due to post-training optimization

journals.sagepub.com/doi/10.1177/...

Sage Journals: Discover world-class research Subscription and open access journals from Sage, the world's leading independent academic publisher.

New paper in Social Science Computer Review 🚨

We conducted two experiments to understand the effects of reading AI summaries, focusing on historical events 📜🤖👩‍💻

We found that AI improved factual recall, possibly due to post-training optimization

journals.sagepub.com/doi/10.1177/...

This generally seems like a bad idea - particularly paying Qualtrics to do it.

However, there are various applications of synthetic data, or "silicon sampling", that are worth exploring. We discuss this in the context of the recent SMR special issue on gen AI: journals.sagepub.com/doi/abs/10.1...

New paper in Nature Human Behaviour.

I use a conjoint experiment to test multimodal large language models (MLLMs) for context-sensitive content moderation and compare with human subjects. Methodologically, this demonstrates how social science techniques can enhance AI auditing. 💻🤖💬

Thanks, Rohan. Looking forward to catching up in Toronto in the spring!

Multimodal large language models can make context-sensitive hate speech evaluations aligned with human judgement - Nature Human Behaviour This study examines how multimodal large language models evaluate hate speech. Larger models can make context-sensitive decisions aligned with human judgement. However, pervasive demographic and lexic...

The paper is out now: nature.com/articles/s41...

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. But these systems still make mistakes, and context can cut both ways, eliminating some biases while enabling others. Human oversight remains essential if deployed for moderation.

Additionally, some models are overtly biased and are particularly sensitive to visual identity cues (AI-generated profile pictures). This demonstrates how different data modalities lead to varying levels of algorithmic bias.

When considering the identity of the author, some MLLMs make context-sensitive judgments comparable to human subjects. e.g., less likely to flag Black users for using reclaimed slurs, a common false positive. But the results also reveal less normative decisions regarding so-called "reverse racism".

I find that MLLMs follow a consistent hierarchy of offensive language to humans and show similarities across other attributes. There is heterogeneity across models, particularly the smallest open-weights versions.

New paper in Nature Human Behaviour.

I use a conjoint experiment to test multimodal large language models (MLLMs) for context-sensitive content moderation and compare with human subjects. Methodologically, this demonstrates how social science techniques can enhance AI auditing. 💻🤖💬

Multimodal large language models can make context-sensitive hate speech evaluations aligned with human judgement - Nature Human Behaviour This study examines how multimodal large language models evaluate hate speech. Larger models can make context-sensitive decisions aligned with human judgement. However, pervasive demographic and lexical biases remain, and visual identity cues may amplify disparities.

Article by @thomasdavidson.bsky.social on multimodal LLMs and hate speech: larger models aligned with human judgment, but pervasive demographic and lexical biases remain, and visual identity cues may amplify disparities.

new paper by Sean Westwood:

With current technology, it is impossible to tell whether survey respondents are real or bots. Among other things, makes it easy for bad actors to manipulate outcomes. No good news here for the future of online-based survey research

I'm recruiting multiple PhD students for Fall 2026 in Computer Science at @hopkinsengineer.bsky.social 🍂

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 at kristinagligoric.com & help spread the word!

Sky News does a great job here of showing that right wing voices get outsized amplification on X. The why of it is complicated (we have upcoming work on this). But the fact if it is undeniable. The platform is also full of clickbaity bullshit. Not unrelated.

Happy to announce the UTM Sociology Speaker Series for 2025-2026 @utm-research.bsky.social @uoftsociology.bsky.social

🇫🇷 We are hiring 🇫🇷

Assistant or Associate Professor Position in Computational Sociology @crestsociology.bsky.social @ipparis.bsky.social

Details here (please RT)
www.shorturl.at/E57le

There is one week left to apply to join us at Rutgers! We're hiring an Assistant Professor in Computational Sociology as part of a cluster of new hires in data science and AI.

Applications are due next Wednesday, 10/15.