New paper w/ UK AISI:
Millions of people now use AI to help them write and communicate. In three experiments (14k participants, 3m+ human ratings) we show that AI writing assistance systematically distorts writer personas – their perceived beliefs, personality, and identity.
🧵
Our results highlight the fragility of persona applications and single-round evaluations, and our protocol provides a method for identifying failures. Code and data available at github.com/peluz/persis...
📢 New paper accepted at @eaclmeeting.bsky.social
2026:
Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions
with
@mhedderich.bsky.social
@amodarressi.bsky.social
Hinrich Schuetze
& Benjamin Roth.
Preprint: arxiv.org/abs/2512.12775
Persona-conditioned LLMs are considered for education, healthcare, and simulation purposes, but many evaluations are single-turn. We ask: what happens when personas must be sustained over long dialogues (100+ rounds)?
Key findings:
1) Persona fidelity degrades as conversations progress---particularly in goal-oriented dialogues.
2) Persona-assigned LLMs consistently underperform no-persona baselines in instruction-following tasks.
3) As fidelity degrades, models revert to their no-persona behavior.
We propose a protocol that combines evaluation datasets with persona dialogue prefixes to measure the effect of conversation length on model behavior.
We then use it to measure the impact of length on:
🎭 Persona Fidelity
✅ Instruction Following
🔒 Safety
Hi all!
We are curating SLAyiNG, a dataset of queer slang. To ensure the quality of the final data, we are asking the community for help with annotation.
Sign up at: docs.google.com/forms/d/e/1F...
If you have further inquiries, feel free to contact either me or @leahirlimann.bsky.social directly 🌈
Going to Rabat for #EACL2026? So are we! 🇲🇦
We are bringing a packed schedule of papers, talks, and workshops.
Check out our lineup below and come say hi! 👋 🧵
#NLProc @eaclmeeting.bsky.social
Paul Röttger
Expert persona prompting -- assigning roles such as expert in math to language models -- is widely used for task improvement. However, prior work shows mixed results on its effectiveness, and does not...
