Sol Messing
@solmg.bsky.social
Social Scientist/Research Prof at NYU CSMaP, formerly Twitter. http://solomonmg.github.io
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
2 PM, VCC West Ballroom B
Chair: @eunjikim.bsky.social. Discussant: @mollyeroberts.bsky.social
September 11, 2025 at 6:26 PM
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
2 PM, VCC West Ballroom B
Chair: @eunjikim.bsky.social. Discussant: @mollyeroberts.bsky.social
And here's the same story appearing shortly after and appearing in Infowars
June 20, 2025 at 2:20 PM
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:
June 20, 2025 at 2:20 PM
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:
June 20, 2025 at 2:20 PM
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
June 20, 2025 at 2:20 PM
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
The challenge is estimating recall, and hence F1.
June 20, 2025 at 2:20 PM
The challenge is estimating recall, and hence F1.
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.
June 20, 2025 at 2:20 PM
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.
Now to the paper: We define narratives as a set of actors, events, and claims thereabout, which generally plays out in a single document. Narrative similarity is the overlap between narratives present in a pair of documents.
June 20, 2025 at 2:20 PM
Now to the paper: We define narratives as a set of actors, events, and claims thereabout, which generally plays out in a single document. Narrative similarity is the overlap between narratives present in a pair of documents.
However, this paper was inspired by something darker—Russian propaganda about "Ukrainian bioweapons." We tracked its spread through state media and U.S. mainstream and fringe news websites, showing how narratives diffuse across media ecosystems.
June 20, 2025 at 2:20 PM
However, this paper was inspired by something darker—Russian propaganda about "Ukrainian bioweapons." We tracked its spread through state media and U.S. mainstream and fringe news websites, showing how narratives diffuse across media ecosystems.
I’m especially excited about this publication because I’ve been thinking about this idea for nearly 20 years, ever since applying to grad school:
June 20, 2025 at 2:20 PM
I’m especially excited about this publication because I’ve been thinking about this idea for nearly 20 years, ever since applying to grad school:
🧵 NEW PAPER - "Quantifying Narrative Similarity Across Languages" is live at SMR. We combine NLP w pairwise LLM evals to track ideas across media ecosystems and languages. We also figured out a way to eval unsupervised probs on widely understood supervised performance metrics.
June 20, 2025 at 2:20 PM
🧵 NEW PAPER - "Quantifying Narrative Similarity Across Languages" is live at SMR. We combine NLP w pairwise LLM evals to track ideas across media ecosystems and languages. We also figured out a way to eval unsupervised probs on widely understood supervised performance metrics.
4. Now guess what?
Musk has 204.9M followers.
Trump Jr has 12.9M followers.
Cory Booker has 4.5M followers.
Rs in this sample also *systematically* post more:https://x.com/SolomonMg/status/1858161536822358214/photo/1
Musk has 204.9M followers.
Trump Jr has 12.9M followers.
Cory Booker has 4.5M followers.
Rs in this sample also *systematically* post more:https://x.com/SolomonMg/status/1858161536822358214/photo/1
November 17, 2024 at 3:15 PM
4. Now guess what?
Musk has 204.9M followers.
Trump Jr has 12.9M followers.
Cory Booker has 4.5M followers.
Rs in this sample also *systematically* post more:https://x.com/SolomonMg/status/1858161536822358214/photo/1
Musk has 204.9M followers.
Trump Jr has 12.9M followers.
Cory Booker has 4.5M followers.
Rs in this sample also *systematically* post more:https://x.com/SolomonMg/status/1858161536822358214/photo/1
2. It's hard to see but ALL accounts in the study saw increased visibility during that timeframe (as the authors note). There were only 5 Dem and 5 Rep accounts, which should raise eyebrows, but let's ignore that for now.
November 17, 2024 at 3:15 PM
2. It's hard to see but ALL accounts in the study saw increased visibility during that timeframe (as the authors note). There were only 5 Dem and 5 Rep accounts, which should raise eyebrows, but let's ignore that for now.
1. It's very likely that *something* in Twitter/X's home-timeline ranking system (HTRS) changed in July 2024.
November 17, 2024 at 3:15 PM
1. It's very likely that *something* in Twitter/X's home-timeline ranking system (HTRS) changed in July 2024.
🧵The study everyone here is talking about does NOT provide evidence that Twitter/X pushed a pro-Republican home timeline ranking change in July 2023.
The cascading, multiplicative effects of ranking changes likely explain the effect--details below
The cascading, multiplicative effects of ranking changes likely explain the effect--details below
November 17, 2024 at 3:15 PM
🧵The study everyone here is talking about does NOT provide evidence that Twitter/X pushed a pro-Republican home timeline ranking change in July 2023.
The cascading, multiplicative effects of ranking changes likely explain the effect--details below
The cascading, multiplicative effects of ranking changes likely explain the effect--details below
Racial attitudes are growing more correlated w party [red box = most of data]. Racial appeals should be less likely to be effective in persuading undecided voters assuming this continues.
The few w higher racial prejudice are already voting R by a huge margin.
HT @dhopkins1776.bsky.social
The few w higher racial prejudice are already voting R by a huge margin.
HT @dhopkins1776.bsky.social
November 17, 2024 at 1:34 PM
Racial attitudes are growing more correlated w party [red box = most of data]. Racial appeals should be less likely to be effective in persuading undecided voters assuming this continues.
The few w higher racial prejudice are already voting R by a huge margin.
HT @dhopkins1776.bsky.social
The few w higher racial prejudice are already voting R by a huge margin.
HT @dhopkins1776.bsky.social
Now it's true that Trump support is highly correlated with various measures of racism, especially the reactionary "anti-anti-racism" measure.
November 15, 2024 at 3:18 PM
Now it's true that Trump support is highly correlated with various measures of racism, especially the reactionary "anti-anti-racism" measure.
Interestingly, if I’d weighted the house election results to 50-50 with the presidential election, the forecast was for a decisive Trump victory. (I weighted things toward the presidential returns at 70-30). Here’s what that looks like:
x.com/SolomonMg/st...
x.com/SolomonMg/st...
November 15, 2024 at 2:52 PM
Interestingly, if I’d weighted the house election results to 50-50 with the presidential election, the forecast was for a decisive Trump victory. (I weighted things toward the presidential returns at 70-30). Here’s what that looks like:
x.com/SolomonMg/st...
x.com/SolomonMg/st...
That “Time for a Change” model failed to predict Trump in 2024. This could be due to using GDP growth instead of cumulative inflation. The public doesn’t “feel” GDP growth in the same way they feel the impact of high prices and high real interest rates.
November 15, 2024 at 2:52 PM
That “Time for a Change” model failed to predict Trump in 2024. This could be due to using GDP growth instead of cumulative inflation. The public doesn’t “feel” GDP growth in the same way they feel the impact of high prices and high real interest rates.
In fact, this is how political scientist Alan Abramowitz successfully predicted Obama’s victory in 2008 (actual was 53%).
pbs.twimg.com/media/Gb4RCO...
pbs.twimg.com/media/Gb4RCO...
November 15, 2024 at 2:52 PM
In fact, this is how political scientist Alan Abramowitz successfully predicted Obama’s victory in 2008 (actual was 53%).
pbs.twimg.com/media/Gb4RCO...
pbs.twimg.com/media/Gb4RCO...
Aside: so far they seem to be wrong—Trumps promised tax cuts and higher tariffs are expected to increase inflation, which means lenders are keeping real interest rates high.
November 15, 2024 at 2:52 PM
Aside: so far they seem to be wrong—Trumps promised tax cuts and higher tariffs are expected to increase inflation, which means lenders are keeping real interest rates high.
Did these people really think Trump’s policies would make their lives better? It seems so. Work from @janzilinsky.bsky.social, Chris Schwarz, Jonathan Nagler, @jatucker.bsky.social, @csmapnyu.org, showed that more voters said Trump’s policies would make their lives better.
November 15, 2024 at 2:52 PM
Did these people really think Trump’s policies would make their lives better? It seems so. Work from @janzilinsky.bsky.social, Chris Schwarz, Jonathan Nagler, @jatucker.bsky.social, @csmapnyu.org, showed that more voters said Trump’s policies would make their lives better.
And almost 3/4 of exit poll respondents who said inflation had hurt them voted for Trump.
November 15, 2024 at 2:52 PM
And almost 3/4 of exit poll respondents who said inflation had hurt them voted for Trump.
Exhibit C: Voters making under 100K went from supporting Biden in 2020 by >10% to supporting Trump. Among those making over 100K, there was an 11 point swing *toward Harris.*
November 15, 2024 at 2:52 PM
Exhibit C: Voters making under 100K went from supporting Biden in 2020 by >10% to supporting Trump. Among those making over 100K, there was an 11 point swing *toward Harris.*
Evidence? Exhibit A: Black voters and women shifted TOWARD Trump in 2024 relative to 2020. What? How can that be? Read on…
November 15, 2024 at 2:52 PM
Evidence? Exhibit A: Black voters and women shifted TOWARD Trump in 2024 relative to 2020. What? How can that be? Read on…