Laura Bronner

🚨My first preprint is out on @socarxiv.bsky.social!
How do AI-generated content labels shape what people see as authentic on social media — and do labels have unintended side effects? osf.io/preprints/so...

A thread 🧵

Going forward, we're planning to add additional annotations and analyses to help us better understand trends in online commenting behavior. Let me know if there's anything you're particularly interested in!

(2) What else can we learn about online behavior? We find that a small minority - 5% of users - is responsible for 78% of the hateful comments submitted to these platforms. But they're active in general, with lots of innocuous posts as well.
www.public-discourse.org/en/articles/...

By tracking how discourse on major Swiss platforms is evolving, we can answer the question: Is it true that online discourse is getting worse?

In this data, in this period, on these platforms, on these measures: no. There are spikes in toxicity and hate, but no overall trend - up or down.

(1) Basic description: what's happening? How many comments are submitted to these platforms every week, and what share is toxic/hateful? @nicolaiberk.bsky.social built classifiers to measure toxicity and hate, and scores are debiased every week using LLMs to avoid drift in the models.

Is online discourse getting worse?

To bring actual data to bear on this question, we launched the Public Discourse Indicator, a dashboard tracking online comments submitted to several major Swiss newspapers. Our aim here is twofold:
www.public-discourse.org/en/public-di...

✖️Who produces most online hate speech and how effective is counterspeech?

➡️ @gloriagennaro.bsky.social et al. find that hate speech is concentrated among a few users and that counterspeech on X mostly fails to curb prolific offenders www.cambridge.org/core/journal... #FirstView

I assume not? Would also be interesting

These average solve times don't seem to account for out-selection as the puzzles get harder... Mondays/Tuesdays probably include a bunch of people who never do the rest of the week. It'd be interesting to show the average number of solvers per day, for context.

This is so cool, congrats!!

Seminars and lectures at IAS Welcome to IAS public lectures and seminars.

Join us on Thursday, 18 Sept, at 14:30 CET for the International Roundtable on Computational Social Science with Laura Bronner @laurabronner.bsky.social 🔹Tackling harmful online comments on news platforms: three field experiments 🔹More info: liu.se/en/article/s...

Counterspeech encouraging users to adopt the perspective of minority groups reduces hate speech and its amplification on social media - Scientific Reports Scientific Reports - Counterspeech encouraging users to adopt the perspective of minority groups reduces hate speech and its amplification on social media

📢 In a new online field experiment, we find that #counterspeech with perspective-based messages can reduce online hate speech, and its 🌟amplification🌟

with a large research team across #ETHZurich @ipz.bsky.social @uclspp.bsky.social and beyond

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

Wow, thank you for helping to keep this alive. I get so many notes from people using this in their coursework.

It seems that Disney never really knew what to do with 538 (Nate's take below), which feels like a real missed opportunity. I hope other outlets will take up the mantle, hire those laid off yesterday, and really invest in data and rigor in journalism - which is more important now than ever.

This is a huge loss. I feel awful for everyone who was laid off, but I'm also just really sad that ABC News didn't appreciate the special blend of reporting chops, data skills, talent, and kindness they managed to amass. I think the *wrong* lesson to draw is that this blend isn't profitable -

6) That said, 538 was also special for journalism - exemplified, perhaps, by the decision to have someone on staff whose entire purpose was to slow stuff down: work through code, question analyses, and be annoying about causal claims. They cared about getting stuff right, even if it took longer.

5) At its best, journalism at 538 blended qualitative (reporting, deep understanding of the substance) with quantitative (data, advanced methods). Academics often think that good research is only done in academia. I think a lot of fantastic research is done in journalism.

4) Understanding what kind of effort gets you 90% of the way to answering something - and whether those 90% are enough to say something meaningful - is something I should remind myself of over and over. Academia spends an inordinate amount of time on those last 10%. Often, it's not worth it.

3) Good data is everything, and understanding data sources and their downsides is crucial for anyone who works with data. So much work went into collecting and auditing the data 538 used - it's a resource for people across (and beyond!) journalism.
bsky.app/profile/base...

2) At the same time, data-centric doesn't necessarily mean complex. Many of the most interesting analyses (e.g. differences in means) are simple; the difficulty is in understanding the data and the substance well enough to ensure those analyses and comparisons are meaningful.

1) Quantitative description and comparison is key to understanding what's going on. Some of 538's most important work was descriptive (e.g. poll trackers, geographic mapping), and much of their contribution to political journalism was in normalizing a much more data-centric type of reporting.

The end of 538 is a huge shame - both for the incredible people who worked there, and for political and data journalism as a whole.

I was lucky enough to work beside them for a few years and want to say a bit about what I think was so valuable that I hope doesn't vanish from the media landscape: 🧵

Out now open access at
@ajpseditor.bsky.social. 194 potential exclusion-restriction violations for studies using weather as an instrumental variable onlinelibrary.wiley.com/doi/full/10....

Interested in how people talk to each other online? Care about causal inference and/or NLP? Want to design and implement field experiments?

Come do a PhD with Dominik Hangartner, me, and a bunch of awesome people at IPL in Zurich:
jobs.ethz.ch/job/view/JOP...

Interested in how people talk to each other online? Care about causal inference and/or NLP? Want to design and implement field experiments?

Come do a PhD with Dominik Hangartner, me, and a bunch of awesome people at IPL in Zurich:
jobs.ethz.ch/job/view/JOP...

What was US2020? The US 2020 Facebook and Instagram Election Project (US2020) was a sprawling project with dozens of people involved at various levels.

The US2020 project was not independent (and that's good) open.substack.com/pub/drewdimm...

This is consistent with stricter pre-publication review, though, right? It's just more of a focus on preventing rather than correcting errors. (Errors in articles/charts were routinely caught by attentive readers, so the incentive to avoid them was strong.)

Clearly 538 values the trade-off differently (or did while I was there), which I think is interesting. I wonder if this partially depends on where the blame for an error goes - the authors or the publication.

It would have required a more in-depth check than this, since our code did produce our results/figures - it's just that the code contained an error. And while correcting the record is great, preventing such errors from being published is also important!
bsky.app/profile/adam...

To put it in personal terms: I would much have preferred the error invalidating my Voting at 16 article to be caught in pre-publication code review, even though the resulting replication-and-extension collaboration probably couldn't have gone any better.
bsky.app/profile/laur...