A paper asks when AI interpretation has earned trust
The risk is false closure, where a clean answer looks settled too soon. Before people treat an AI reading as reliable, they need to check the sources, the rules for judgment, the allowed claims, the error plan, and who must fix problems.
What the source reports
Deyu Jing's arXiv paper is a concept paper, not a benchmark. A benchmark is a test used to compare AI systems. The paper says local checks for facts, citations, coverage, and structure do not prove that a reading is established. Here, reading means a claim about texts, culture, or history. The paper names three ideas. Interpretive appearance means an answer looks finished, but its process is not clear. An evaluation contract states the sources, task, criteria, allowed claims, and failure rules. Standing substitution means a local pass gets treated as a stronger public claim. Jing calls for delayed closure. Public review should name sources and versions, explain evidence roles, state failure rules, revise the contract, and assign responsibility.
Original source
- Title
- When Does an Interpretation Count as Established? The Formation, Evaluation, and Responsibility of Interpretation in Generative AI
- Author
- Deyu Jing
- Publication
- arXiv
- Date
- Friday, September 4, 2026