Sol Messing

IF ANYONE BUILDSIT, EVERYONE DIES WHY SUPERHUMAN AI WOULD KILLUS ALL ELIEZER YUDKOWSKY & NATE SOARES

Next up on my reading list.

…I am already regretting this choice.

How common are “survey professionals” - people who take dozens of online surveys for pay - across online panels, and do they harm data quality?

Our paper, FirstView at @politicalanalysis.bsky.social, tackles this question using browsing data from three U.S. samples (Facebook, YouGov, and Lucid):

7/ Kernel Approximation Ideal Point Poisson Factorization - scalable Bayesian ideological estimation from billions of observations, applied to 134 million TikTok comments to map ideology on the platform. @matiaspiqueras.bsky.social

6/ Propaganda or Parity? testing whether TikTok amplifies pro-China content, using LLM classifiers and longitudinal engagement/moderation data. @kengchichang.bsky.social, @mollyeroberts.bsky.social, H Barnehl

5/ To Be or Not to Be on TikTok: a rare activation experiment, recruiting users to start TikTok and measuring causal effects on attitudes, knowledge and well-being.
K Rutherford, @tiagoventura.bsky.social

4/ Scrolling Through Hate: mapping hate speech on TikTok across time, place, topic, plus experiments testing moderation responsiveness.
responsiveness.
@karstendonnay.bsky.social, @fabriziogilardi.bsky.social , @gloriagennaro.bsky.social, @dhangartner.bsky.social

3/ First: The Political Supply of TikTok: political content spreads faster than entertainment, and a small set of creators dominates reach. @benguinaudeau.bsky.social, K Rutherford, @jatucker.bsky.social

2/ Short-form video platforms (TikTok, Reels, Shorts) are reshaping political comms: vertical video, personalized feeds, huge reach. But opaque data access makes them hard to study.

TODAY Aug 28 - "Politics in 60 Seconds: Short-Form Video, TikTok, and Political Communication" at #APSA2025 -

2 PM, VCC West Ballroom B

Chair: @eunjikim.bsky.social. Discussant: @mollyeroberts.bsky.social

Spoiler alert: the answer is “no”

Maybe let’s not dismantle the tenure system while admin and athletic budgets balloon

How exactly did Grok go full 'MechaHitler?' After Grok took a hard turn toward antisemitic earlier this week, many are probably left wondering how something like that could even happen.

New from me about Grok's very bad week, with insight from @solmg.bsky.social. www.engadget.com/ai/how-exact...

Reposting @hwaight.bsky.social ‘s thread on this bsky.app/profile/hwai...

Here's a link to the paper! solomonmg.github.io/pdf/Quantify...

Limitations exist—measuring narrative similarity doesn't alone prove "diffusion." Contextual and temporal analyses remain essential for robust conclusions about propaganda or any information dynamics.

Authorship Analysis:
Current methods often employ BERT for authorship attribution. However, larger, modern LLMs (like GPT-4o) remain under-explored for this task. There's untapped potential here waiting to be studied.

Science of Science:
Tracking idea origins in scientific literature traditionally uses topic models or exact text reuse, often missing important conceptual linkages. Our method could clarify how ideas propagate through academic communities.

Information Reuse:
The study of content recycling ("memetracking") relies heavily on exact text matches. Using our approach could identify deeper connections—tracing the subtle evolution and spread of ideas.

Our method has potential beyond narrative similarity. Here are some potential applications:

Plagiarism Detection:
Exact-text matching often misses subtle, paraphrased copying. Our approach could vastly improve recall, catching nuanced cases traditional methods miss.

And here's the same story appearing shortly after and appearing in Infowars

What does this look like in practice? Here's an article in Sputnik alleging a false-flag operation by the US:

Now it's important to note that often matching narratives represent humdrum coverage of the same real-world developments:

And here's what we actually found: low quality news outlets with lower journalistic standards are more likely to print narratives appearing in Russian state media outlets

We use purposive sampling at various decision boundaries to oversample positive cases to generate labeled training and validation data sets. This allows us to estimate recall and thus F1!

The challenge is estimating recall, and hence F1.

But wait how did we get supervised metrics for an unsupervised problem? We build on Grimmer
& @garyking90.bsky.social
(2011) & Mozer et al (2020) generating within & between-cluster measures of validity, but we estimate [out-of-sample] precision, recall, & F1.

Narrative similarity is also not the same as topical overlap, which is too inclusive of documents containing different narratives on the same topic.

Why? Because narrative similarity is *not the same* as text similarity (used in cheating detection software), which often relies on exact text features and often misses narratives that are phrased differently.

Validation for something like this is crucial. We benchmarked against existing methods—exact text reuse, topic modeling, semantic role labeling. Our LLM-based method significantly outperformed others on precision and recall.