arXiv paper uses Conrad's Typhoon to probe AI self-consumption and steering
Researchers describe an audiovisual installation in which a large language model, seeded with Joseph Conrad's novella, feeds on its own output while activation steering shapes the drift.
In plain words
A team built an art installation where a language model starts from Conrad's Typhoon and then keeps rewriting its own text. Over time the writing grows strange, wandering into the model's instructions and incoherent passages. The authors used a steering technique, which nudges the model's internal activity, to control how the story unfolds.
What was announced
The installation, described in a new arXiv abstract, prompts a large language model seeded with Conrad's original language to consume its own output endlessly. The authors chose the novella, in which Captain MacWhirr steers the Nan-Shan into a catastrophic storm, as a lens on the AI moment: the book examines what happens when a phenomenon outgrows the frameworks available to understand it. As generation proceeds, they report, the text becomes increasingly alien, with narrative deviations, digressions into the model's own instructions and thinking, and long stretches of incoherent rumination.
To manage this drift, the team applied activation steering methods, including a novel adaptive technique they introduce, to control the rhythm of the narrative and the model's digressions over long generations. They conclude that the steering's simultaneous efficacy and unpredictability illustrates the tension between the perceived interpretability of AI and the reality of a system that remains outside our comprehension and control. The evidence is limited to the paper's abstract, so the full methods, evaluation details and results are not visible here.
What it does not show
- Coverage is limited to the paper's abstract; the full article's methods, quantitative results and evaluation are not available in the supplied passages.
- The steering is described only qualitatively; no measurements of its effectiveness or unpredictability are provided in the excerpt.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.