The abstract of a new arXiv paper (2605.17193) opens with a question that has persisted since the nineteenth century: can a machine originate anything? It invokes Ada Lovelace's famous denial that an engine could "originate anything" and Alan Turing's later framing of whether machines can amplify ideas. The paper appears to use this historical debate as a lens for evaluating contemporary multi-LLM systems.

The paper's title announces its central finding: "robust semantic collapse." In multi-LLM systems, where several models are combined or queried together, outputs tend to converge on the same underlying meaning even when their surface wording differs. The term "robust" suggests this collapse is not incidental but persists across different configurations or prompts.

If the finding holds, it complicates the notion that running multiple language models in parallel yields genuine diversity of thought. Instead, the systems may be recirculating a shared semantic core. The paper thus ties a technical observation about LLM ensembles back to the older philosophical question of whether machines can truly originate anything.