Albert Wendsjö
LightNews LightNews
albertwendsjo.bsky.social
Albert Wendsjö
@albertwendsjo.bsky.social
210 followers 150 following 0 posts
PhD Student in Political Science at the University of Gothenburg. awendsjo.com
Posts Replies Media Videos
Reposted by Albert Wendsjö
taniseceron.bsky.social
📣 New Preprint!
Have you ever wondered what the political content in LLM's training data is? What are the political opinions expressed? What is the proportion of left- vs right-leaning documents in the pre- and post-training data? Do they correlate with the political biases reflected in models?
September 29, 2025 at 2:54 PM
Reposted by Albert Wendsjö
joachimbaumann.bsky.social
🚨 New paper alert 🚨 Using LLMs as data annotators, you can produce any scientific result you want. We call this **LLM Hacking**.

Paper: arxiv.org/pdf/2509.08825
We present our new preprint titled "Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation". We quantify LLM hacking risk through systematic replication of 37 diverse computational social science annotation tasks. For these tasks, we use a combined set of 2,361 realistic hypotheses that researchers might test using these annotations. Then, we collect 13 million LLM annotations across plausible LLM configurations. These annotations feed into 1.4 million regressions testing the hypotheses. For a hypothesis with no true effect (ground truth $p > 0.05$), different LLM configurations yield conflicting conclusions. Checkmarks indicate correct statistical conclusions matching ground truth; crosses indicate LLM hacking -- incorrect conclusions due to annotation errors. Across all experiments, LLM hacking occurs in 31-50\% of cases even with highly capable models. Since minor configuration changes can flip scientific conclusions, from correct to incorrect, LLM hacking can be exploited to present anything as statistically significant.
September 12, 2025 at 10:33 AM
Reposted by Albert Wendsjö
carlmc.bsky.social
🚨 Very excited that our paper on *Rulers on the Road* has been cond. accepted at the AJPS @ajpseditor.bsky.social. We analyze emperors' strategies of itinerant rule in the Holy Roman Empire 919-1519. Fun working with @claranw.bsky.social, @andrejkokkonen.bsky.social & Jørgen Møller shorturl.at/Spm7z
Itineraries of 25 Emperors of the Holy Roman Empire, 919 to 1519 Itinerant rule, rule exercised through traveling, was a common, yet insufficiently researched pre-modern form of governance. Studying the determinants of ruler itineraries in the Holy Roman Empire AD 919-1519, we argue that rulers focused on monitoring `marginal' elites. Powerful rulers could count on family members and thus targeted unrelated local elites. Weak emperors had to monitor their less loyal relatives and left unrelated nobles unvisited. We reconstruct emperors' itineraries from 72'665 dated and geolocated documents and measure territorial control by their relatives. Exploiting the weakening of imperial power through the Great Interregnum (1250-1273), we find that strong, pre-1250 emperors frequented areas controlled by their relatives relatively less. In contrast, family control increased visits post-1273. Causal identification rests on the discontinuous reduction of emperors' power through the Great Interregnum and differences in family relations between subsequent emperors. The results show strategic itinerant rule as an important but understudied form of governance.
April 29, 2025 at 3:43 PM