Thomas Stephan Juzek

Research & publications

I am interested in the human language system, the foundations of language technology, human-AI interaction, and the societal impact of language technology.

My current research focuses on AI’s linguistic fingerprint and AI-associated language change: how model language behaviour diverges from human language behaviour, and how the two interact. This work connects to broader questions in model alignment, including how models come to reflect, amplify, or depart from human expectations, and what their language use reveals about their inner workings.

Below is a selection of recent projects, grouped by theme.

In brief

What is “AI-associated language change”?

“AI-associated language change” refers to the observation that there are systematic word-preference shifts in human language use, and the words involved also happen to be the ones that AI overuses, and the changes in human behaviour started to kick in right after the release of ChatGPT in 2022. It is thus plausible to associate these changes with large language model output. The phrasing is deliberately neutral about causation. Thomas Stephan Juzek’s research group at Florida State University works on this question.

Is AI changing human language?

There is a measurable association for sure: the words that models like ChatGPT overuse have risen sharply in Scientific English and have seen a still pretty remarkable increase in English news writing, as well as in semi-spontaneous spoken English. Such increases are quite something historically and they are difficult to explain without AI. At the same time, causality is always tricky to show. For example, for news English and spoken language, a plausible alternative hypothesis is that these changes would have happened without AI, too, but AI is accelerating these trends.

Why does ChatGPT overuse words like “delve”?

Part of it goes back to a model training stage called (Reinforcement) Learning from Human Feedback: human annotators ever so slightly but systematically prefer text containing certain words, and preference training then amplifies exactly those words. At the same time, there are still a lot of open questions. Models first learn from enormous amounts of human-written text during a stage called pre-training, and we still do not fully understand how tendencies acquired there interact with later preference training.

Does this happen outside English?

Yes; this is a recent study of mine. When you let AI do news writing, the same concepts surface among AI-overused words in 24 of 34 languages; emphasize-type verbs were the clearest case. And in a second analysis of human news writing, AI-preferred words show post-ChatGPT rises in 26 of 34.

An AI global fingerprint

Across languages, ChatGPT leaves a kind of cross-lingual fingerprint: it overuses the same concepts in dozens of the world’s languages. And exactly those overused words have risen markedly in human usage since ChatGPT’s release in 2022.

AI-Associated Lexical Shifts Across 34 Languages: Cross-Lingual Convergence and Diachronic Uptake in News Writing

Juzek, T. S. · preprint, 2026

A cross-lingual “AI register”: emphasize-type verbs surface among AI-overused words in 24 of 34 languages; AI-preferred words rise +15.1% in post-ChatGPT news vs −4.5% for matched baselines.

arXiv GitHub

The same trend appears in speech. Testing (semi-)spontaneous spoken English, we find that AI-overused words have increased markedly from before to after ChatGPT’s release.

Model Misalignment and Language Change: Traces of AI-Associated Language in Unscripted Spoken English

Anderson, B., Galpin, R. & Juzek, T. S. · AIES 2025

The first peer-reviewed evidence of AI-associated vocabulary appearing in unscripted spoken English.

Proceedings arXiv GitHub

The why

But why does ChatGPT overuse these words? For Scientific English, the literature has documented striking lexical shifts and conjectured that AI is the cause. We strengthen that link (the words that have been spiking are also the ones AI overuses), and then examine the mechanisms, finding that learning from human preferences (RLHF) plays an important role. Two papers explore this.

Why Does ChatGPT “Delve” So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models

Juzek, T. S. & Ward, Z. B. · COLING 2025

Why LLMs overuse words like delve, underscore, and intricate: identifying focal words whose rise in scientific abstracts tracks LLM use.

Proceedings arXiv GitHub

Word Overuse and Alignment in Large Language Models: The Influence of Learning from Human Feedback

Juzek, T. S. & Ward, Z. B. · BIAS 2025 Workshop (ECML PKDD)

Links AI lexical overuse to learning from human feedback: annotators systematically prefer text containing certain words, and preference training amplifies them.

arXiv GitHub

Model diagnostics

Much of the literature on AI and language still depends on hand-curated word lists at some step. We automate the diagnostics for model language behaviour and alignment, with no manual curation required.

Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models

Juzek, T. S., Ming, X. & Hernandez, J. A. · LREC 2026, 6116–6131

A curation-free pipeline (Lexical Alignment Score plus a triangulated preference shift) that derives AI-overused word inventories without manual lists: the foundational diagnostic for this research line.

Proceedings GitHub

Also recent

Karolina Rudnicka and I have a new preprint, Beyond “AI Language”: The case for the idiolectal nature of LLM output (2026): across two model cohorts, each model shows a distinct linguistic profile of its own, closer to an idiolect than a single shared register.

This work has been covered by The New York Times, The Guardian, The Economist, and many other outlets worldwide: see the press coverage of this research.

For the complete, up-to-date list with citation counts (including earlier work on experimental syntax, acceptability, and corpora), see Google Scholar.