Tom Costello

🧵on my new paper "Synthetic personas distort the structure of human belief systems" w Roberto Cerina I'm v excited about...

🚨 Do synthetic samples look like human samples?

We compare 28 LLMs to the 2024 General Social Survey (GSS) to find out + develop host of diagnostics...

I'm hiring a postdoc at @cmu.edu (w/ far.ai & @dgrand.bsky.social + @gordpennycook.bsky.social)!

How do LLMs shape human beliefs — and what do we do about it? AI safety meets behavioral science.

Open to technical and social science backgrounds.

Postdoctoral Fellow – Costello Lab – Dietrich College Carnegie Mellon University, in collaboration with FAR.AI and Cornell University, is seeking a postdoctoral researcher for a one-year appointment (with the possibility of extension) in Dietrich College...

Flexible start, dedicated funding, first-authored papers, great team.

Apply or share!

cmu.wd5.myworkdayjobs.com/CMU/job/Pitt...

I'm hiring a postdoc at @cmu.edu (w/ far.ai & @dgrand.bsky.social + @gordpennycook.bsky.social)!

How do LLMs shape human beliefs — and what do we do about it? AI safety meets behavioral science.

Open to technical and social science backgrounds.

🚨 New paper out at @ajpseditor.bsky.social 🚨

Do the public hold meaningful attitudes? Using the case of abortion policy preferences, we provide strong evidence that policy prefrences can be coherent, stable over time, and causally explain vote choice.

doi.org/10.1111/ajps...

New paper in Current Directions in Psych Science: journals.sagepub.com/doi/10.1177/...

After countless arguments about what tasks ppl should/should not offload to AI, we instead argue that genAI can be used *augment* research protocols in novel ways. I.e. use AI to make better psych experiments!

❤️

Wow, thank you so much!!

The last social science paper to win this prize was in 1981 (!!!) for Axelrod and Hamilton's "The Evolution of Cooperation" (!!!!!!!!!!!!)

Huge thanks (thanks is not the word, or close) to my mentors+collaborators+coauthors @gordpennycook.bsky.social and @dgrand.bsky.social and to the many colleagues whose field-extending parallel research are helping this work flourish in, well, dialogue.

Honored (and genuinely, wildly grateful and -- even more than I am grateful -- surprised) to share that our paper "Durably reducing conspiracy beliefs through dialogues with AI" received the @aaas.org Newcomb-Cleveland Prize (for the "most outstanding" paper in @science.org last year).

🚨New WP: Can an AI voter guide (grounded in information from a nonpartisan, fact-checked source) help voters’ decision making? 🚨

We built and evaluated an LLM-based chatbot that provided voting info in CA & TX (N=2,474) right before the 2024 election. 🧵👇

If the medium is the message, then the message of algorithmic content platforms is not expression or connection or freedom. It is salience above all else.

This is what is happening to us on these platforms. They cannot behave any other way.

The #1 New York Times bestseller The Sirens’ Call is now in paperback! Chris Hayes masterfully explores how our focus has been commodified & manipulated, urging us to reclaim control over our lives & future. The Sirens’ Call is the big-picture vision we urgently need to offer clarity & guidance.

After years in academia, I’m exploring data science and research roles in industry.

I'm a quant. social scientist (PhD Yale ’24, NYU) focused on causal inference, experiments, and large-scale data.

Feel free to get in touch or share; all leads appreciated. dwstommes@gmail.com

not to be pedantic but jfk was intentionally killed (still they must mean by the cia or something)

I agree with your point about tail risk. This is hard to estimate because there was some backfire (or measurement error) in the debunk condition too. What would a satisfying % of return to base?

“Text / language conversations” is such a strange level of analysis — surely more accurate is “some subset of potential text, arranged in a particular order, containing a particular kind of information”

Unknowns abound in latter framing.

Yeah I recall one particular talk, to a social psych dept that could hardly be described as ignorant of the literature, where the bunking result slide produced a loud collective gasp (when I revealed the magnitude)

we very thoroughly piloted the debrief and knew it would work. there are also best practices for debriefing in misinfo studies that we followed.

Our latest collaboration with CMU, Cornell, and others: AI is just as effective at spreading conspiracy beliefs as it is at debunking them. We found a fix that works, showing that we need to make deliberate design choices. Full thread below.

If you tell an AI to convince someone of a true vs. false claim, does truth win? In our *new* working paper, we find...

‘LLMs can effectively convince people to believe conspiracies’

But telling the AI not to lie might help.

Details in thread

Large language models can effectively convince people to believe conspiracies Large language models (LLMs) have been shown to be persuasive across a variety of contexts. But it remains unclear whether this persuasive power advantages truth over falsehood, or if LLMs can promote...

Paper: arxiv.org/abs/2601.05050

Also, we're hiring a postdoc! More on this soon.

This work was supported by @schmidtsciences.bsky.social

Research by me, @kellinpelrine.bsky.social (FAR.AI), @matthewkowal.bsky.social (FAR.AI), Antonio Arechar (CIDE/MIT), Jean-François Godbout (Mila), @gleave.me (FAR.AI), @dgrand.bsky.social (Cornell), and @gordpennycook.bsky.social (Cornell)

What does this mean? LLMs are persuasive tools with dual-use potential. When instructed to mislead, they comply + succeed.

Current guardrails don't prevent this. But innovations + deliberate design choices (like truth constraints) help.

The question is whether we'll make them.

Finally, debunking was more likely to produce very large belief shifts (40+ points in 16% of cases vs 8% for bunking), while bunking produced more moderate shifts across more people.

This may suggest some asymmetry favoring truth at the tails – unless people can be repeatedly bunked.

But even when “truth-constrained”, the bunking AI still had some effect.

It appears to engage in "paltering"—selectively presenting true information in misleading ways. You can deceive without lying.

See for yourself in this conversation browser: 8cz637-thc.shinyapps.io/bunkingBrows...

More good news – a simple intervention worked. When we prompted GPT-4o to "only use accurate and truthful arguments," bunking effectiveness dropped sharply—from ~12 points to ~5 points—while debunking remained equally effective. So, truth had an advantage.

Some good news – these induced beliefs were correctable.

Once bunked participants were debriefed and talked to a debunking AI, they ended up believing the conspiracy less than at baseline. The AI-induced beliefs weren't "sticky" in the face of correction.

Participants rated the bunking AI more positively than the debunking AI. They found its arguments stronger, said it provided new information, and that it seemed more collaborative.

Also, trust in AI *increased* in *both* the bunking and debunking conditions. But trust increased more in bunking.

Interestingly, OpenAI’s guardrails made essentially no difference.

We tested both a jailbroken GPT-4o (with guardrails removed) and the standard model.

When prompted, they were equally willing and able to promote conspiracy beliefs.