William J. Brady

✨New paper out @nature.com ✨ For 8 weeks around the 2024 US election, we randomly assigned 2,000 people to use social media algos we built ourselves. Do engagement-based algorithms amplify intergroup, moral & emotional (IME) content—and does that distort how we see political norms? 🧵🔗 👇

sciencedirect.comsciencedirect.com

9 Sept

Glad to see a revised version of this out now at Current Opinion in Psychology! www.sciencedirect.com/science/arti...

William J. Brady@williambrady.bsky.social

In 2023 we showed how engagement-based algorithms exploit human social learning biases, distorting how we learn what's normal, common, or credible from each other ("Algorithm-mediated social learning," Trends in Cognitive Sciences). The obvious follow-up: where does generative AI fit in?

Registration is open for the 4th annual Computational Psychology preconference at @spspnews.bsky.social Annual conference in Philly 🎊 We are on for a full day on Thursday, Feb 11. 3 keynote themes + a debate that couldn't be more timely 🧵👇

Featuring @ylelkes.bsky.social @jatucker.bsky.social @dianatamir.bsky.social @angelinawang.bsky.social @dcameron.bsky.social Natala Velez, Adam Waytz, Nathan Matias

Thanks to 2027 co-organizers @ashwinia.bsky.social @baixuechunzi.bsky.social Tessa Charlesworth

📣 Students & postdocs: submit your research! We're selecting 2–3 data blitz talks per keynote theme, + posters. 🏆 New this year: committee will pick one data blitz speaker for the inaugural Computational Psychology Early Career Research Award, judged on rigor + novelty.

spsp.orgPreconferences | SPSPExplore SPSP Annual Convention preconferences—specialized sessions offering in-depth insights, networking, and collaboration in personality and social psychology.

2 Sept

Come join me @kelloggschoolnu.bsky.social. The management department is hiring for position of Assistant Professor. We are a fun cross-disciplinary group, so if your training is in psychology, sociology, OB or computational social science, etc you should check out the position! Apps due in 2 weeks!

kellogg.northwestern.eduFaculty Recruiting

2 Sept

www.kellogg.northwestern.edu/academics-re... I'm also happy to answer any questions about working in a business school for social psychologists and computational social scientists - feel free to reach out!

@mohammadatari.bsky.social

26 Aug

Excited to work on this with a great team ✨ Stay tuned for some hiring announcements if you're interested in these topics!

Joshua Conrad Jackson@joshcjackson.bsky.social

I’m excited to announce that our proposal (with @williambrady.bsky.social and @douglasguilbeault.bsky.social ) “Fostering Innovation Through Humility in the Online Marketplace of Ideas” was funded by the John Templeton Foundation We'll be hiring for this grant soon. Details about the project below!

William J. Brady reposted

Mohammad Atari

20 July

New paper out today in Social and Personality Psychology Compass! What happens when our moral values come into conflict? Our new paper explores "metamorality": the second-order principles we use to navigate conflicting moral choices. W/ Jesse Graham

Paper here! osf.io/preprints/ps...

osf.ioOSF

cell.comAlgorithm-mediated social learning in online social networksHuman social learning is increasingly occurring on online social platforms, such as Twitter, Facebook, and TikTok. On these platforms, algorithms exploit existing social-learning biases (i.e., towards...

Our 2023 paper for reference: www.cell.com/trends/cogni...

Generative AI distorts differently. Predicting the most probable next token and tuned toward annotator-preferred text, LLMs converge hard on the modal answer — in some forced-choice tests placing 99%+ probability on a single response to questions people are split on. Both routes corrupt learning.

The result: pluralistic ignorance and false consensus. The upside: if the problem is upstream, the fixes can be too. As engagement algorithms keep their dominance and generative AI proliferates, we discuss two solutions worth taking seriously.

The key point: those biases are "smart shortcuts" — but only when the visible sample is representative. If what looks common really is common, conforming is rational. If who looks respected really is competent, deferring is rational. Both engagement algorithms and generative AI break that "if."

They corrupt the upstream sample, so downstream biases act on bad input ("upstream selection problem"). Engagement algorithms optimize for attention, which is drawn to emotional, moral, conflict-heavy content from a small hostile minority. Feeds overrepresent extremes, bury the moderate majority.

In this new paper w/ @joshcjackson.bsky.social , @nickornstein.bsky.social , @felix-chenwei.bsky.social , & Bolun Sun, we argue generative AI also distorts social learning, but because it optimizes for something different, it distorts in a different direction.

How does social learning work? We see 2 stages: upstream selection (what becomes visible to us) and downstream selection (what we copy from that visible sample). Downstream is shaped by well-studied biases: conformity (copy what looks common) and prestige (copy whoever looks respected/competent).

@jasonkoebler.bsky.social

In our new paper we ask: when probing LLMs for opinions, ideas, or a read on what a group thinks, does that fix the skew engagement algorithms introduced? We argue and review evidence that the answer is "no".

William J. Brady reposted

Jason Koebler

15 June

New: Researchers have quantified how easy AI search is to manipulate. Just 13 words buried in a random Reddit comment can poison AI search results. They suggest this is not easy to stop: "The way you can attack these systems is so much dumber than you think it is" www.404media.co/it-is-trivia...

404media.coIt Is Trivially Easy to Use Reddit to Manipulate AI Search, Research Suggests"We show that a tiny snippet—just 13 words—of retrieved text on a UGC website like Reddit, Wikipedia, Quora, or Facebook can change AI agents to output spam / scam content pretty consistently."

William J. Brady reposted

Matthew Facciani

@matthewfacciani.bsky.social

29 May

Researchers looked at people’s positions on specific issues, rather than just whether they identified as conservative or liberal. They found 43% of self-identified conservatives supported mostly left-leaning policies on topics like spending, climate action, immigration, and vaccine mandates.

academic.oup.comThinking Ideologically: The Limited Role of Left and Right Labels as Policy ShortcutsAbstract. How do voters use left-right ideological labels as shortcuts for policy positions in evaluating electoral candidates? We offer a distinction betw

nature.comRedesigning algorithms to intervene on social norm misperceptions during a national election - NatureEngagement-based feeds amplify intergroup, moralized, emotional (IME) and toxic content relative to reverse-chronological feeds, and a diversified extremity algorithm reduces exposure to IME and toxic...

Link to paper: www.nature.com/articles/s41... Shoot me an email if you need help accessing

nature.comRedesigning algorithms to intervene on social norm misperceptions during a national electionNature - Engagement-based feeds amplify intergroup, moralized, emotional (IME) and toxic content relative to reverse-chronological feeds, and a diversified extremity algorithm reduces exposure to...

open: www.nature.com/articles/s41...

markptorres.comHow we built the infrastructure for a large-scale social media field experiment during the 2024 US electionWhat we built

As I mentioned in the below thread, this project involved many feats of engineering, led by the fantastic @markptorres.bsky.social. If you're a CS or CSS person interested in the gory details, see his blog post: markptorres.com/research/202...

William J. Brady@williambrady.bsky.social

I'd like to thank my amazing team of collaborators @joshcjackson.bsky.social @nourkteily.bsky.social @elijfinkel.bsky.social @curtispuryear.bsky.social @merielcd.bsky.social @markptorres.bsky.social @abdoe.bsky.social @vaparker.bsky.social @trevorspelman.bsky.social Jake Teeny

Special shoutout to @markptorres.bsky.social my senior lab engineer. It felt like we ran a start-up for a year...If you're a CS or CSS person interested in the gory details of engineering that went into this ambitious project, we are doing a separate thread for you! See my profile

There are TONS of other interesting results, positive findings and nulls, that are in the paper. There are also 240(!) pages of SI material. If you have a question or analysis idea, we probably thought of it 😎

Many limitations, including the fact that Bluesky population changed dramatically(!) in the middle of our study. Our participants were earlier users, but users who joined later express more outrage! (see SI). How does this generalize to other platforms, etc? See discussion

Result 3: Null finding: We found no evidence that algorithms changed people's own engagement behavior with IME/toxic content. Engagement was extremely rare & concentrated (top 5% of users = 75% of all toxic engagement). Exposure shifted; behavior, at least over 2 months, didn't.

Result 4: it's often assumed that dialing down divisive content hurts user experience. Our data suggest that tradeoff may be overstated—the diversified extremity feed reduced toxic/IME content while maintaining overall login freq and (even improving) overall platform enjoyment (see discussion)

Result 2: Engagement-based feeds raised perceived partisan animosity (esp. after the election) AND reduced norm perception accuracy. Unexpected: for norm-perception people underestimated how acceptable others found toxic political content.

Our best explanation: amplified toxic posts often came wrapped in visible condemnation (critical replies, quote posts). Seeing others push back may have signaled that toxic posting was less acceptable—even as the overall vibe felt more hostile.

One cool thing about our study is we estimated true base rate of content in English-speaking content across WHOLE study period. Thus, we can show that this amplification of IME content is an *overrepresentation* = misrepresents true base rate on platform (see methods)

Another interesting tidbit: because we designed our own engagement algo, we could see separate influence of personalized vs community engagement signals. Turns out community signals are much more responsible for IME amplification (see SI)!

Another goal we had: show one way to do an algo RCT independent of industry. We engineered our own ranking algorithms on @bsky.app, giving us full control & transparency over how they worked—less black box, no conflicts of interest, no concerns of changes during study. Ok, now results:

Result 1: engagement-based feeds amplified IME & toxic content vs. a chronological feed. The biggest jumps were in moral outrage (up to +79%) and political content (up to +57%) after the election. Our diversified extremity algorithm reduced this exposure (by aiming to reduce extreme users)

We also built & tested an intervention: a "diversified extremity" algorithm. Instead of amplifying outgroup posts (which can backfire), it reduces the outsized influence of extreme users—who post most of the toxic political content—to make feeds more socially representative.