PS: Massive thanks to my colleague Dr. Alessandro Miani for his invaluable contribution to this project.

Read more about GERMA here: www.degruyter.com/document/doi...
The corpus is freely available here: osf.io/3bthj/ Check it out and let us know how you use it. 6/6

GERMA is part of PRODEMINFO, an ERC-funded project studying misinformation in Europe and its links to differing ontologies of truth. It contributes to developing multilingual corpora of news from unreliable sources such as IRMA, our dataset for Italian news: aclanthology.org/2023.eacl-ma... 5/6

3/ Website-level metadata: GERMA goes beyond articles by offering data on bias, credibility, transparency, and traffic patterns—allowing researchers to study digital communities and misinformation spread. 4/6

2/ Rich text-level data: GERMA includes full texts, topics, publication dates, and keywords for linguistic studies. It also provides 90+ LIWC categories (e.g., emotional tone, cognitive processes) for psychological research. 3/6

GERMA is not the first of its kind—so what makes it unique? Three key features:

1/ Scale: GERMA contains 230K+ articles from unreliable domains. While individual article credibility still requires verification, this approach allows for a broader collection of texts for large-scale analysis. 2/6

PS: a massive thanks goes to my co-authors @aroyehuns.bsky.social , @janalasser.bsky.social, @almogsi.bsky.social, @dgarcia.eu , @lewan.bsky.social.

You can read the full story here: www.nature.com/articles/s41...

While belief-speaking is linked to greater affective polarization and lower quality of information shared (see our previous paper: www.nature.com/articles/s41...), a more evidence-based communication style can foster healthier, more constructive online discourse. (8/8)

Key takeaway: Political leaders shape the tone of online conversations. In an era where parties use entire social media platforms as their own broadcasting stations, holding them accountable for how they communicate is crucial. (7/8)

Belief-speaking tweets were also associated with more affectively polarized language in replies. Fact-speaking ones, on the other hand, showed a negative correlation with affective polarization. However, we didn't replicate this correlation under experimental conditions. (6/8)

More importantly, this causal link was confirmed by a preregistered experiment with controlled settings. Even with synthetic political messages, participants aligned their responses with the honesty framing of the initiating tweet. (5/8)

Our study establishes a strong correlation, but causality is tricky. However, by tracking users who replied to both fact- and belief-based tweets from the same politician, we found evidence of linguistic adaptation—suggesting a causal link. (4/8)

Interestingly, the contagion is stronger when the politician and the user share the same party affiliation. This means that political leaders play a crucial role in shaping intra-partisan discourse on social media. (3/8)

We analyzed +10K tweets from U.S. politicians and found that replies tend to mirror the honesty style of the original tweet. Fact-speaking prompts fact-based replies. Belief-speaking prompts belief-driven replies. This holds across party lines and user partisanship. (2/8)