As large language models (LLMs) become primary generators of online articles, those articles are increasingly fed back as inputs to subsequent models. This feedback loop links each individual response to the pool of information available for later generation. We focus on a crucial distinction: an article may avoid explicitly endorsing a false claim while still preserving the concrete details that enable the claim to be reproduced. To capture this, we propose a framework that ties claim content, expressed stance, and downstream reuse, explaining why errors can persist selectively under a continual mix of accurate and inaccurate information.
We conduct a controlled community‑writing experiment that varies whether published articles are returned as inputs, aligning external supply, input capacity, and generation volume. The study spans twelve factual targets and four generation models. Results show that recirculation raises the prevalence of complete false claims by 7.29 percentage points in the final round, while explicit endorsement drops by 8.33 points. The magnitude of these effects differs across facts and models.
We then apply paired interventions that abstract eligible error details within fixed histories, preserving correct content and the surrounding controversy. Evaluation indicates that error reproduction remains largely unchanged across most targets, with only a development‑quotation target showing a localized reduction.
These findings link persistent errors to the information available throughout the writing process and motivate evaluating not only what models affirm but also what their publications supply to future writers.
Review