iiinova — where your digital self keeps evolving

The broader lesson is that improving science is not only about collecting better data. It is also about making sure the claims built on that data accurately reflect the evidence underneath them, and that those limits are preserved when research is communicated and summarized

100

Together, the findings show why causal overclaiming matters beyond a single paper. Scientific claims travel to researchers, journalists, policymakers, the public, and AI summaries, where claims beyond the evidence can shape what people believe the research established!

100

In other words, how LLMs are instructed to summarize scientific evidence can affect whether those evidentiary limits are preserved.

100

But the researchers also found a way to reduce this problem! When LLMs were explicitly instructed to respect the limits of the research design and avoid unsupported claims, causal overclaiming decreased substantially.

100

That result points to a challenge for science communication. Prompts intended to make research easier to understand or more useful can inadvertently make its conclusions sound stronger than the evidence warrants

100

The researchers further examined what happens when LLMs summarize this research. Several common prompting approaches increased causal over-claiming, including asking models to simplify a study to an eighth-grade reading level or explain its practical implications

100

Using correlational wording alone did not prevent causal inference. Readers often still inferred causality. A brief note on the study design and what it could and could not establish reduced this over-inference, although the effect was small

But does causal wording actually change what readers think the evidence shows? The researchers conducted experiments with 1,105 participants. They found that readers exposed to causal wording were more likely to believe that a study provided causal evidence.

100

How common is it? The researchers examined 194,631 cross-sectional studies, which generally provide evidence of association rather than causation. They found causal claims in about 46% of titles or abstracts, with causal overclaiming increasing roughly threefold from 2000 to 2024

100

The authors call this broader problem “Narrative License” when claims about a study extend beyond what was actually demonstrated. This paper focuses on one form of it, causal claims based on correlational evidence

100

Quantifying the prevalence and impact of overreaching causal claims in social science - Nature Human Behaviour

What happens when scientific claims go beyond what a study can establish? In new Nature Human Behaviour research, Calvin, Timothy, Neil, Grace, and Duncan find causal claims in 46% of 194,631 cross-sectional studies and show that causal wording makes readers more likely to infer causation

nature.com

An analysis of 194,631 social science articles found causal overclaiming in 46% of cases, increasing over time. This language shaped reader interpretations; AI-generated summaries often amplified over...

125

CSSLab at Penn @csspenn.bsky.social · 25/08/2026

Beyond his own research, he has been an important part of the intellectual life of the lab. His thoughtful feedback, questions, and engagement have helped many of our projects move forward. Congratulations again, Samar! We’re excited to see where you take this work next 🚀

000

CSSLab at Penn @csspenn.bsky.social · 25/08/2026

Congratulations, Dr. Samar Haider! 🎓🖥️ @samarhaider.bsky.social Samar successfully defended his dissertation, "A Computational Approach to News Production", studying how computational methods can help us systematically understand what news organizations choose to cover and how they cover it.

101

This work is part of the broader NOMAD effort to expand access to human mobility data and develop methods, tools, and guidance for analyzing it reliably!! 🛰️Learn more about NOMAD nomad.seas.upenn.edu

000

The results show that different approaches can achieve similar accuracy when tuned to individual trajectories, but vary in robustness across a dataset. Smaller spatial thresholds can split or miss stops, while larger ones can merge neighboring stops

100

They built a synthetic environment with known movements and stops, then sparsely sampled trajectories to mimic real world location data. This allowed them to test eight stop detection algorithms and measure when stops are correctly identified, missed, split, or merged

100

Measures like "stops🛑" are building blocks for studying human mobility. But how reliably can we identify them in sparse GPS data?🛰️📍 At the CSS Showcase, Paco Barreras, Andrés Mondragón, and Caroline Chen shared work on the robustness of stop detection in sparse location data

CSSLab at Penn @csspenn.bsky.social · 14/08/2026

What have we been working on at the CSS Lab? 👀 A quick look inside our Spring Showcase, with lightning talks, posters, new tools, and works in progress across computational social science. A few highlights from the day 🎥 And of course, more to come!

043

CSSLab at Penn @csspenn.bsky.social · 07/08/2026

Retrieval-augmented generation (RAG) substantially improved how closely LLM responses matched human survey responses, outweighing differences in model size. Even so, demographic gaps persisted, especially for older, independent, and minority respondents.

010

CSSLab at Penn @csspenn.bsky.social · 07/08/2026

Using nearly 19,000 human survey responses collected across 40 waves during the 2024 U.S. presidential election, Elliot Pickens(@elliot-p.bsky.social), David Rothschild, Jenny Wang, and Jonas Mikhaeil compared LLM-generated and human survey responses. #IC2S2

121

Using 120,000 quotations from more than 30,000 news articles as the unit of analysis, the study provides large-scale empirical evidence of differences in political representation in U.S. news quotations during the 2024 presidential election.

000

Samar Haider @samarhaider.bsky.social presented Whose Voice Matters? Measuring Political Representation in News Quotations, with Chris Callison-Burch and Duncan Watts @duncanjwatts.bsky.social! #IC2S2

143

Analyzing more than 600,000 U.S. news articles with LLM-assisted content analysis, the study examines whether national news coverage disproportionately emphasizes some crimes over others, amplifies fear around crime, and frames punishment as the primary response.

000

Amir Tohidi @atohidi.bsky.social presented at #IC2S2 Measuring "Copaganda": An Analysis of How U.S. National News Covers Crime, with Baird Howland, Billy Pierce, Duncan Watts @duncanjwatts.bsky.social, and David Rothschild @davmicrot.bsky.social.

123

The work, coauthored with Michael Kearns, Aaron Roth, and Duncan Watts @duncanjwatts.bsky.social, examines how people evaluate competing notions of algorithmic fairness and how the presentation of trade-offs shapes perceptions of what is fair.

000

At #AOM2026, Bethany Hsiao presented Fairness Preferences in Algorithmic Decision-Making as part of the symposium AI in Organizations: Challenges to Face and Opportunities to Embrace. @aomconnect.bsky.social

101

Robin Na presented work with Duncan Watts @duncanjwatts.bsky.social and Abdullah Almaatouq on whether grounding LLM predictions in scientific literature improves their ability to predict behavioral experiments, finding that literature conditioning rarely improved prediction accuracy.

020

Homa @homahm.bsky.social presented work with Pavel Savgira, Elisa Kreiss, and Amir Ghasemian @amir-ghasemian.bsky.social, on how LLM-generated summaries change the political lean of news articles by examining when summarization moderates political lean and when it preserves more partisan content.

120

Josh Nguyen@joshnguyen.bsky.social presented work with Upasana, Samar, Bryan, Neil, and Duncan on using LLM-based agent simulations to study the consequences of unbalanced online news consumption. It examines how political attitudes vary with the partisan content agents are systematically exposed to

Timothy Dörr and Baird Howland presented their work on US and UK news media coverage of Gaza, leveraging GPT-5 to extract structured descriptions of violence. The study examines volume of coverage, emotional weight of violence descriptions, and naming of perpetrators.

Where could you find the most CSS Lab folks in one place at #IC2S2? The poster session! Here are a few of the projects our lab members presented today.

111

At today's #IC2S2 Methods and Measurement session, Linnea Gandhi @linneagandhi.bsky.social presented her study with Elizabeth Tipton and Duncan Watts, which tests whether research syntheses can predict the results of new studies and highlights the challenges of producing generalizable knowledge.

032

CSSLab at Penn @csspenn.bsky.social · 30/07/2026

Do falsehoods persuade more than selective facts? Jenny Allen presented Falsehoods Offer No Persuasive Advantage Over Selective Facts: Evidence from Real-World Political Misinformation at today's #IC2S2 Plenary Lightning Talks with David Rothschild, Duncan Watts, Amir Tohidi, and Samar Haider!

022

Baird and Grace presented their work at #IC2S2, introducing a systematic framework for comparing official White House communications across administrations through the lens of authoritarian-populism! www.linkedin.com/posts/ic2s2_...

linkedin.com

Afternoon parallel: Baird Howland explores shifts in White House media in the Political Polarization and Discourse session: | IC2S2

Afternoon parallel: Baird Howland explores shifts in White House media in the Political Polarization and Discourse session:

050

CSSLab at Penn @csspenn.bsky.social · 29/07/2026

Friday, July 31 Mansfield(210) | Track H | 10:45–12:15 • Can LLMs Approximate Public Opinion? Validating Synthetic Survey Responses Against 40 Waves of Real Survey Data by Elliot Pickens (@elliot-p.bsky.social), David Rothschild (@davmicrot.bsky.social), Jenny Wang, and Jonas Mikhaeil

010

Building on her recent work, she highlighted best practices for transparent, reproducible, and ethical TikTok research in computational social science.

001

Gayoung (@gayoung-jeon.bsky.social) led a hands-on tutorial on comprehensive TikTok data collection for computational social science at #IC2S2'26 Participants learned practical methods for collecting and analyzing TikTok data. Great work on a successful tutorial!

101

Friday, July 31 Room 764 | 2:15 PM – 3:45 PM • Whose Voice Matters? Measuring Political Representation in News Quotations Samar Haider (@samarhaider.bsky.social), Chris Callison-Burch, and Duncan Watts (@duncanjwatts.bsky.social) #CSSLab #IC2S2 #ComputationalSocialScience #Burlington

Friday, July 31 Sugar (400) | Track E | 2:15 PM – 3:45 PM • Measuring "Copaganda": An Analysis of How U.S. National News Covers Crime by Amir Tohidi (@atohidi.bsky.social), Baird Howland , Billy Pierce, Duncan Watts(@duncanjwatts.bsky.social), and David Rothschild(@davmicrot.bsky.social)

Friday, July 31 Silver (401) | Track D | 2:15 PM – 3:45 PM • Beyond Total Survey Error: How Reporting and Specification Errors Can Mislead Stakeholders by Hannah Cha and David Rothschild (@davmicrot.bsky.social )

100

Friday, July 31 Silver (401) | Track D | 2:15 PM – 3:45 PM • Agentic Incidence and Data Quality in Online Samples by David Rothschild (@davmicrot.bsky.social) and Andrew Gordon Chaired by Elliot Pickens(@elliot-p.bsky.social)

100

Friday, July 31 Williams (403) | Track C | 2:15 PM – 3:45 PM • Seeking Help, Facing Harm: Auditing TikTok's Mental Health Recommendations by Pooriya Jamie, Amir Ghasemian (@amir-ghasemian.bsky.social) and Homa Hosseinmardi (@homahm.bsky.social)

Friday, July 31 Williams (403) | Track C | 2:15 PM – 3:45 PM • Understanding Media Diet and App Usage Across Multiple Short-Form Video Platforms by Thomas Ma, Francesco Corso, Jennifer Allen, and Dean Eckles

100

Thursday, July 30 Aiken 110 | Track D | 2:15 PM – 3:45 PM • How Well Does the Literature on Behavioral Interventions Generalize? An Empirical Test by Linnea Gandhi (@linneagandhi.bsky.social), Elizabeth Tipton, and Duncan Watts(@duncanjwatts.bsky.social)

120

Thursday, July 30 Grand Maple Ballroom | 9:00 AM – 9:45 AM • Falsehoods Offer No Persuasive Advantage Over Selective Facts: Evidence from Real-World Political Misinformation by Jennifer Allen, David Rothschild, Duncan Watts, Amir Tohidi, and Samar Haider

100