iiinova — where your digital self keeps evolving

Simon Munzert

@simonsaysnothin.bsky.social

682 followers 563 following 125 posts

Professor of Data Science and Public Policy | Director @ Hertie School Data Science Lab | Elections, Public Opinion, Data

Posts Replies Media

Simon Munzert @simonsaysnothin.bsky.social · 01/07/2026

For instance, in "The Quiet record", Elena Dreyer follows a European radar satellite passing over Ukraine every 12 days. Over 4 years of war, its signals built a record of Russia's strikes on Ukraine's power grid. A stunning piece of satellite journalism. elenaivadreyer.github.io/final-project/

Screenshot of a map of Ukraine, highlighting Russian air strikes targeting energy infrastructure.

120

Simon Munzert @simonsaysnothin.bsky.social · 01/07/2026

Stop! 🛑 Scroll no further - this is the coolest thing you'll see today. Presenting The Data Dispatch: data journalism made by my students at the Hertie School 👇 data-journalism-26.github.io

1256

Simon Munzert @simonsaysnothin.bsky.social · 21/05/2026

Opa, wenn er zu Heinos Rock-Platte wippt

ZEIT-Artikel über Ulf Poschardt „Ich sehe mich in der Tradition des Punk“

100

Simon Munzert @simonsaysnothin.bsky.social · 18/12/2025

Same vibes

Image from Spinal Tap where he says it goes up to eleven

010

Simon Munzert @simonsaysnothin.bsky.social · 31/10/2025

This week, the 2025 edition of the Lancet Countdown Report on Health and Climate Change was released — and I was able to contribute a small part again (look for indicator 5.2 in our version of the #ShowYourStripes chart, that’s me!). 1/4

Chart 1 of 2 from Lancet Countdown 2025 providing time-series information on climate and health indicators reported in the paper. Lots of unfavorable, a few favorable developments.

Chart 2 of 2 from Lancet Countdown 2025 providing time-series information on climate and health indicators reported in the paper. Lots of unfavorable, a few favorable developments.

150

Simon Munzert @simonsaysnothin.bsky.social · 12/05/2025

Main finding #2: Group markers drive over-moderation. Words like "muslim", "gay", or "jews" make mis-classifying non-hate speech as hate speech more likely.

Table providing results from a SHAP Values analysis, showing which speech tokens contribute to misclassification decisions.

100

Simon Munzert @simonsaysnothin.bsky.social · 12/05/2025

Main finding #1: Performance varies wildly across moderation services, datasets, and metrics, but some of the failure rates are astonishing (FPR & FNR > 75% on balanced samples for some implicit and explicit speech!).

Table reporting some key findings from the paper - hate speech classification performance statistics across moderation services and datasets.

100

Simon Munzert @simonsaysnothin.bsky.social · 12/05/2025

Our paper investigates how commercial content moderation APIs handle group-targeted hate speech. This is a black-box audit (we have no idea how the models look exactly) of 5 major APIs with five million queries based on four datasets.

Info chart describing the design of the study. Figure title: Our black-box audit framework to evaluate commercial content moderation APIs.

100

Simon Munzert @simonsaysnothin.bsky.social · 12/05/2025

Have you ever used a content moderation API, such as Perspective API or OpenAI's Moderation API, for your research or to inform moderation decisions? Well, they might not have given you what you think they would.

Bill Murray sitting on a bed, somehow lost. Scene from movie Lost in Translation. Title says "Lost in Moderation"

100