DIGHUM lectures: LLMs When Left Alone

Speaker: Stefan Szeider, TU Wien, AT
Moderator: Julia Neidhardt, TU Wien, AT

Today, autonomous AI agents are so widely deployed that millions of people interact with them daily. These include coding assistants, chatbots, and systems that run for hours without human oversight. A natural question to ask is whether something is going on inside these systems beyond sophisticated pattern matching. The question is easy to dismiss but hard to settle. The standard dismissals (anthropomorphism, reductionism, training bias) each carry force, but none is conclusive. In this talk, we follow an empirical approach. Instead of speculating, we placed frontier language models in minimal autonomous loops with no task and no user. We observe that the behavioral patterns that emerged are surprisingly stable and model-specific. The patterns proved stable enough to identify the source model with high accuracy. We further examined whether AI self-reports of internal states can be trusted. We find that a “placebo tool” that does nothing reliably shifts what models say about themselves. These findings raise questions that go beyond computer science and concern society at large. These questions will only grow more pressing as AI agents become more autonomous.

About the Series

A roughly bi-weekly seminar offers presentations and panels from worldwide thought leaders. It is typically held on Tuesday afternoons at 17:00 CET. The bidt and the TU Wien are cooperation partners for the DIGHUM lecture series. Digital Humanism deals with the complex relationship between man and machine. It acknowledges the potential of Informatics and IT. At the same time, it points to related apparent threats such as privacy violations, ethical concerns with AI, automation, loss of jobs, and the ongoing monopolisation on the Web. For this reason, a new initiative — DIGHUM lectures — started with regular online events to discuss the different aspects of Digital Humanism. We will have one or more speakers on a specific topic followed by a discussion, or panel discussions, depending on the topic and speakers. The exact dates will be announced at least two weeks before.

Participate via Zoom
Zoom | meeting: 9638 9928 143, password: 0dzqxqiy

Participate VIA YouTube
Talk will be live-streamed and recorded via YouTube Archive and Resources of Lecture Series Abstracts, Recordings, and Slides from Previous Talks