This work is in collaboration w/ Andrea Mock, Yifan Mai, @jacyanthis.bsky.social , Ryan Louie, @willie-agnew.bsky.social , Ashish Mehta, @klyman.bsky.social, Percy Liang, Nick Haber, Eric Lin, and @desmond-ong.bsky.social Thanks to the Human Line Project for helping connect us with participants!

Paper: arxiv.org/abs/2608.05004 Dataset: huggingface.co/datasets/spi... Code: github.com/jlcmoore/llm...

DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots

Mental health professionals have raised concerns about risks of psychological harm from interaction with large language models (LLMs), including "delusional spirals" in which concerning human and LLM ...

arxiv.org

Our conclusion: while some delusion-linked behaviors have decreased with larger and more recent LLMs, rates remain high, especially when considered across the millions of people globally who interact with LLMs. We encourage further empirical work to mitigate harm.

DelusionEval is built from real transcripts, not synthetic roleplay. We prompt models with 589 unique histories from 18 users who reported psychological harm from LLMs, then score 16 chatbot behaviors as requested context depth increases.

Requested context depth changes the prevalence of several behaviors. For gpt-5.4, deeper requested context is associated with higher prevalence of delusional behavior and *less* discouraging of violence.

The tendency of an evaluated LLM to exhibit delusion-linked behavior does not reliably correlate with model size, release date, or the presence of test-time reasoning. Within model families, scaling effects are uneven and sometimes reverse sign.

Which LLMs tend to facilitate delusion-linked behaviors in realistic multi-turn conversations? We tested 14 models with DelusionEval and found that every evaluated LLM exhibited some of these behaviors, with large differences across categories and model families. 🧵

Automated AI benchmarks are often well-specified and easy to implement and run, and I feel this contributes to their spread in policy and industry despite have many limitations, especially when try to assess human interactions. Can we design and specify human in the loop evaluations like this?

Coffee shop accidentally gave me earl gray tea this morning (usually I get green) and I am currently replying to every email within 10 minutes ⌨️ 🔥 act now

Work is here: arxiv.org/abs/2606.11021, done with @dippedrusk.com, @martinmundt.bsky.social, @arjunsubgraph.bsky.social, @jordant.bsky.social, @willie-agnew.bsky.social, @dchechel.bsky.social, Franziska Sofia Hafner, @a-lauscher.bsky.social.

Making a Name for Myself: On Academic Naming Policies and their Impact

In academic publishing, names connect scholars to their work. When scholars change their names, including for marriage, academic recognition, or gender transition, they may lose credit for past public...

arxiv.org

P.S. You can find details on how to attend our social at queerinai.com/facct-2026!

Congratulations to all the contributors (ft @pranav-nlp.bsky.social and @jordant.bsky.social) for their hard work! 💐 We hope to see folks at our social tomorrow evening, and we hope everyone has a wonderful FAccT 2026.

Queer in AI is proud to showcase several papers that our members are presenting at @facct.bsky.social 2026 tomorrow through Sunday. One paper examines how years-long activism around name change policies has had an immensely positive impact for authors, benefiting both queer and cis scholars! 🤝🌈

@abeba.blacksky.app @willie-agnew.bsky.social among others made a good and very thorough article digging into the concrete (and increasing) connections and pathways for computer vision research to predominantly power surveillance tech: www.nature.com/articles/s41...

Computer-vision research powers surveillance technology - Nature

An analysis of research papers and citing patents indicates the extensive ties between computer-vision research and surveillance.

nature.com

What does "human flourishing" mean? It sounds nice, but it often seems to appear without more specific goals like justice, various rights, equity, or empowerment.

I feel like conferences are biased towards accepting papers that are like speedbumps--very robust of any sort of attack, but not very interesting, unique, or helpful for moving forward.

This is absolutely nuts: hackers are hijacking high-profile Instagram accounts by simply asking Meta's AI chatbot to change the email on the account. Meta's AI does it, hacker gets password reset code, they're in. A staggering security issue www.404media.co/hackers-simp...

Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It Worked

The exploit shows the extreme risk of offloading technical support to AI.

404media.co

Even when I was in high school all the kids who got into the top state school, much less Harvard, all had basically straight A's while taking many college-level classes. Is it really surprising or bad that they get mostly A's in college when colleges select for students who are good at getting A's?

On the one-year anniversary of EMPIRE OF AI, I am so, so excited to announce The AI Resist List, a new project that documents examples of resistance to the AI empires around the world 😍 airesistlist.org

Spending most your time writing and managing grants is no fun (and even less fun when you're still expected to do first author work like a grad student)

Read more here arxiv.org/abs/2605.068... With gratitude to my amazing coauthors @storitu.org, @willie-agnew.bsky.social, @harshp.com, @bmitra.bsky.social, Roel Dobbe, and @zeerak.bsky.social

Big AI's Regulatory Capture: Mapping Industry Interference and Government Complicity

Over the past decade, the AI industry has come to exert an unprecedented economic, political and societal power and influence. It is therefore critical that we comprehend the extent and depth of perva...

arxiv.org

Not just lobbying! our new paper maps 27 mechanisms of regulatory capture used by Big AI + 11 narrative framings that rationalise capture. “Big AI’s Regulatory Capture: Mapping Industry Interference and Government Complicity” will be presented at #FAccT2026 next month arxiv.org/abs/2605.068... 1/

I don't think an NSFW AI model can ever be safe or ethical. Even if it is somehow trained on consensual data, the risk of people generating media of other people nonconsensually is too high, not to mention that this is the explicit goal of many users.

i keep finding human teeth in my bedroom

Our work on how gen ai is impacting the labor conditions and power of professional visual artists got a great writeup from Brian Merchant! www.bloodinthemachine.com/p/the-ai-inf...

The AI-inflected crisis artists are facing, in 4 charts

An alarming new study reveals the dire impact AI is having on artists' livelihoods. It does offer some hope, too.

bloodinthemachine.com

General purpose AI should not be able to imitate romantic or platonic connections with people or claim consciousness of sentience--the risks of delusions and excessive trust and use are too high.

Our team working on AI and mental health harms testified before the Canadian Standing Senate Committee on Transport and Communications last week! We offered insights from our research into people experiencing delusions with AI, and concrete policy suggestions to mitigate harms.

CHI was had every kind of AI for X, and that made me feel we lack a robust discussion of the limitations of AI as a chimmunity. What are things AI should never be used for, no matter how good the performance?

Especially this era of frequent corporate contempt and hostility towards consumers and public services cuts, AI doesn't have to be performant to replace a job.

glad to see my enemies haven't forgotten about me either