I wrote a guide to red teaming AI tools for safety for @indicator.media. It's a combination of simple frameworks and snippets of worked-through examples built from my experience at Google, running a pro bono clinic at Cornell Tech, and doing this kind of work as a journalist.

How to red team an AI tool for safety

Indicator’s no-code and product-agnostic guide to probing for harmful stuff

indicator.media

Really excited for this! We have an amazing program lined up! 🤩🤩🤩

Stanford Tech Impact and Policy Center@techimpactpolicy.bsky.social · yesterday

The 2026 #TSRConf agenda is now live! Join us at Stanford on October 1–2, 2026 for conversations on platform accountability, #AI governance, online harms, transparency, #DigitalPolicy, and other key #TrustAndSafety issues. Explore the agenda ⤵️ 🔗 bit.ly/3SgzmXP

A little bit late, but this paper has now been published in ACL. Note: more recent (unpublished) analyses suggest agents are getting better at searching the web—which I've been mentioning in my conference presentations this summer.

Large Language Models Require Curated Context for Reliable Political Fact-Checking—Even with Reasoning and Web Search

Matthew R. DeVerna, Kai-Cheng Yang, Harry Yaojun Yan, Filippo Menczer. Findings of the Association for Computational Linguistics: ACL 2026. 2026.

aclanthology.org

Matt DeVerna@matthewdeverna.com · 8mo ago

🚨 New working paper 🚨 Can LLMs with reasoning + web search reliably fact-check political claims? We evaluated 15 models from OpenAI, Google, Meta, and DeepSeek on 6,000+ PolitiFact claims (2007–2024). Short answer: Not reliably—unless you give them curated evidence. arxiv.org/abs/2511.18749

Some initial thoughts, and a complicated mix of feelings. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.

That's a wrap on IC2S2 2026... we'll see you next year in Milan! 😍

Presenting our ACL paper this afternoon at #IC2S2 (Understanding LLMs session @ 2:15) aclanthology.org/2026.finding...

Large Language Models Require Curated Context for Reliable Political Fact-Checking—Even with Reasoning and Web Search

Matthew R. DeVerna, Kai-Cheng Yang, Harry Yaojun Yan, Filippo Menczer. Findings of the Association for Computational Linguistics: ACL 2026. 2026.

aclanthology.org

**Please repost** If you're enjoying Paper Skygest -- our personalized feed of academic content on Bluesky -- we'd appreciate you reposting this! We’ve found that the most effective way for us to reach new users and communities is through users sharing it with their network

New massive persuasion experiment from AISI.

AI systems out-persuade expert humans

Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has...

arxiv.org

We want to hire a new, more techy AI person in the lab. Think mechanistic interpretability, fine tuning, large-scale evals that kind of thing. If you're interested and have Qs, lmk! interfolio.com/181820

Yale Psychology is hiring an Assistant Professor in Clinical Psychology! Review of applications will begin on September 2, 2026. Link: apply.interfolio.com/188806 Please share widely & apply!

We're excited to be hosting CODE@MIT again this fall! We invite a mix of work involving experimentation — methods, cool experiments, industry reports, etc. Look for more info later on this year's invited speakers.

CODE@MIT@codemit.bsky.social · 2w ago

📢 Call for abstracts for the Conference on Digital Experimentation (CODE@MIT), which takes place Nov 13-14. 📨 Submit your best work by *Sept 13th*! ide.mit.edu/events/2026-...

We have re-opened our search for a 3-year AI-research postdoc at NYU's Center for Social Media, AI, and Politics!! The position starts this **September 1st** (or as soon as possible thereafter) and is in person at NYU. If necessary, we will also consider a 1/1/27 start. 1/3

I have a new paper out in @sociologicalsci.bsky.social w/ the briliant @aybukeatalay.bsky.social and @aliaelkattan.com We show that what we pay attention to online is a lot more politically diverse than what public engagement signals would suggest... sociologicalscience.com/articles-v13...

Information Diets Are More Diverse in Attention Than in Engagement

Article: Information Diets Are More Diverse in Attention Than in Engagement | Sociological Science | Posted July 21, 2026

sociologicalscience.com

👇👇👇

International Conference on Computational Social Science@ic2s2.bsky.social · 2w ago

Are you presenting a poster at IC2S2 2026? Here's a handy order form for UVM Print and Mail if you plan to print on site! qualtrics.uvm.edu/jfe/form/SV_...

New in Trends in Cognitive Sciences: misinformation isn't just about false beliefs. Led by Ullrich Ecker & co=authors Emily Spearing & Renee DiResta our piece argues the bigger threat is indirect—the erosion of shared epistemic standards doi.org/10.1016/j.ti... 🧵 1/10

Misinformation as strategy: Epistemic consequences and the undermining of shared truth

Misinformation influences cognition—shaping memory, beliefs, attitudes, reasoning, and decision-making. While intentionally disseminated misinformatio…

doi.org

Text as Data has been a wonderful, long-running, non-archival workshop for empirical research at the intersection of AI and social science (especially work involving text). After a few years off, it will be happening again this year as a one-day event in early October!

Denis@dpeskoff.bsky.social · 3w ago

Do you work across computational methods, social sciences, and the humanities? Submit to Text as Data 2026! 📄 One-page submissions 🔓 Non-archival ⏰ Due August 1 📍 October 5 @UCBerkeley tada2026.org

Text as Data is happening at Berkeley right before COLM. Please share!

Denis@dpeskoff.bsky.social · 3w ago

Do you work across computational methods, social sciences, and the humanities? Submit to Text as Data 2026! 📄 One-page submissions 🔓 Non-archival ⏰ Due August 1 📍 October 5 @UCBerkeley tada2026.org

lol love this

Kate Starbird@katestarbird.bsky.social · 3w ago

In anticipation of Trump's press conference on alleged election interference tomorrow, we're doing an early release of our "Rigged Election" Bingo game. Which worn out tropes will he and his operatives roll out (again) tomorrow? Create your own card to play along. bingo.cip.uw.edu/s/92d3739d-2...

"Whoops! Most arXiv papers contain information never meant to be shared. Passwords, GPS coordinates and private conversations have all been found in the metadata and source files of millions of preprint submissions." www.nature.com/articles/d41...

Whoops! Most arXiv papers contain information never meant to be shared

Passwords, GPS coordinates and private conversations have all been found in the metadata and source files of millions of preprint submissions.

nature.com

As some may have heard me talk about at #ACL2026, I'm excited to share a new preprint on approaches to validation when using LLMs to measure concepts in social science, led by @madesai.bsky.social and @azjacobs.bsky.social !! Paper: arxiv.org/abs/2607.07915

ACL 2026 was fun! We presented our benchmark, which reveals that frontier models struggle to moderate pluralistic communities on Reddit. A fun bonus: creating all the colorful plots for the paper! 🎨 Our paper, code, and benchmark are all available here: github.com/osome-iu/Plu...

See here for @isabelcorpus.bsky.social paper showing that the "write with AI" feature on change.org made petitions longer and their text more complicated. The outcomes? Did not improve. arxiv.org/abs/2511.13949

Introducing AI to an Online Petition Platform Changed Outputs but not Outcomes

The rapid integration of AI writing tools into online platforms raises critical questions about their impact on content production and outcomes. We leverage a unique natural experiment on Change$.$org...

arxiv.org

In case anyone is curious to study how network effects lose critical mass, search inclusion is a good opportunity to update scientific theories of collective action

Rachel Holliday Smith@rachelholliday.bsky.social · 4w ago

I'm been wondering if it would come to this and apparently it has. huge move from cloudflare, which hosts a fifth of all websites on earth.

The Tech Impact and Policy Center is thrilled to have been selected as one of 30 @safeonlinefund.bsky.social grantees. Our project, Co-Designing a Child-Safe LLM, was chosen for its potential as an innovative model to address urgent child protection challenges in the digital age.

Early-bird registration for #TSRConf 2026 closes soon! Join us for interdisciplinary conversations on: ➤ AI governance ➤ platform accountability ➤ online harms ➤ and other key #TrustAndSafety issues Register by July 15 for reduced pricing! 🎟️ stanford.io/3SgrmWl

If you're not sure where to start, the /setup-site skill allows you to simply share screenshots of a site you like, or even a link as inspiration. Then tweak the design however you want via conversation with Claude. You need a website! I hope this helps. 🤓

Matt DeVerna@matthewdeverna.com · 4w ago

One of the first things I learned as a PhD student was that you *need* a website. I built a template site that uses Claude Code to easily design and maintain your personal site—hosted free on GitHub. Clone it, then run the "/setup-site" skill in Claude Code and Claude walks you through set up.

One of the first things I learned as a PhD student was that you *need* a website. I built a template site that uses Claude Code to easily design and maintain your personal site—hosted free on GitHub. Clone it, then run the "/setup-site" skill in Claude Code and Claude walks you through set up.

GitHub - mr-devs/cc-academic-website: Academic website template + Claude Code assisted development.

Academic website template + Claude Code assisted development. - mr-devs/cc-academic-website

github.com

We should pay peer reviewers Paying reviewers led to faster decisions—an average of 5.5 working days vs 38 days for unpaid reviews And review quality improved (judged by handling editors) Reviewers who delivered a good peer review in 4 days were paid $290 www.nature.com/articles/d41...

My upcoming book is absolutely packed to the brim with em dashes—more em dashes than any well-adjusted human being could possibly need or want—and if any of you tries to claim they are the result of some soulless LLM rather than my own unhinged compulsion I will fight you.

news.cornell.edu/stories/2026...

Digital research repository arXiv to start new chapter as nonprofit | Cornell Chronicle

On July 1, the digital research respository arXiv, housed at Cornell Tech, will transition to an independent nonprofit, enabling faster technological development, expanded partnerships and long-term f...

news.cornell.edu