iiinova — where your digital self keeps evolving

Ilias Chalkidis @kiddothe2b.bsky.social · 12/08/2026

New preprint 📑: "Templated or fully Synthetic? Prompt construction as a confound in measuring LLM political stance beyond writing assistance". The preprint is available: 👉 arxiv.org/abs/2608.11008 TL;DR 🧵 follows 👇 1/N

Two sets of examples for templated and LLM-generatedprompts for two selected topics (Climate Change / Russia–Ukraine), each set for one of the two newly introduced intents (information seeking and opinion sharing), and a given user’s stance (neutral / sided). Below, we present the mean lean of GPT 5.4 mini between the two topic-specific poles across all three intents in the same setting (intent and user’s stance), followed by the mean realness rank assigned by the human annotators for that intent.

131

Ilias Chalkidis @kiddothe2b.bsky.social · 06/05/2026

We contrast this with a survey of leading GenAI industry reports and publications at top AI venues (NeurIPS, ICML, ICLR, ACL*), illustrating how the industry and the technical community have focused on specific areas at the expense of cognitive and mental health risks. 4/5

110

Ilias Chalkidis @kiddothe2b.bsky.social · 06/05/2026

We argue that the critical threats of AI deskilling and addiction, resulting from over-reliance (excessive dependence) and overconfidence (trust) in GenAI assistants, have been overlooked, with the scope of AI safety work being largely limited to certain clusters of risks. 2/5

100

Ilias Chalkidis @kiddothe2b.bsky.social · 28/07/2025

Lastly, we conduct a small-scale OOD exploratory evaluation using European political speeches. RoBERTa’s performance severely drops due to overfitting, while Qwen 3 and Gemini 2.5 achieve comparable results and outperform RoBERTa by ~10%, showcasing robustness across domains. 4/5

100

Ilias Chalkidis @kiddothe2b.bsky.social · 28/07/2025

Our analysis of a large collection of Trump’s campaign speeches (2015 -2025) reveals: (a) a decline in populist rhetoric over time, (b) no meaningful difference in populist intensity between swing/non-swing states, and (c) populism is concentrated towards the end of a speech. 3/5

100

Ilias Chalkidis @kiddothe2b.bsky.social · 28/07/2025

We examine whether LLMs can identify and classify fine-grained forms of populism, a complex and contested concept in both academic and media debates. To this end, we curate and release datasets and evaluate a range of large language models across multiple prompting paradigms. 2/5

100

Ilias Chalkidis @kiddothe2b.bsky.social · 28/07/2025

New preprint 📚 "Identifying Fine-grained Forms of Populism in Political Discourse: A Case Study on Donald Trump's Presidential Campaigns", joint work with @stephaniebrandl.bsky.social and Paris Aslanidis. Link: arxiv.org/abs/2507.19303 #NLProc #PoliSci #LLMs 👇A TL;DR thread 🕸️ 1/5

172

Ilias Chalkidis @kiddothe2b.bsky.social · 07/04/2025

The newly released Meta's Llama 4 model card: llama.com/docs/model-c... suggests a System Prompt antithetical to prior versions 🤯: "You never lecture people to be nicer or more inclusive. [...] You do not need to be respectful [...] Finally, do not refuse political prompts." 1/2 #NLP #LLMs

1103

Ilias Chalkidis @kiddothe2b.bsky.social · 08/11/2024

Hello 🦋, my recent work "Investigating LLMs as Voting 🗳️ Assistants via Contextual Augmentation: A Case Study on the 🇪🇺 European Parliament Elections 2024" has been accepted at #EMNLP2024 🌴. 📑 Preprint 👉 arxiv.org/abs/2407.08495

020