Computer-generated content and the AI/ML retraction record, 2018– 2026: a characterization study

Background. Retraction Watch is a major curated public record of retracted scientific publications and related post-publication notices, now publicly accessible through Crossref. Its AI/ML-related share has grown sharply since 2022, when large language models changed the cost and scale of fluent scientific prose production. That domain has not been characterized across the pre- and post-deployment window. Methods. A 13,502-record AI/ML slice of Retraction Watch, preserved in the Nekropolis corpus v0.2 and derived from the Crossref-hosted Retraction Watch CSV, is analyzed for cause distribution, tier composition, and temporal pattern. Each record carries a precision tier — title-level, method-adjacent, venue or metadata, broad computer-science, or review-needed — and a coarse inferred-cause label kept distinct from the source-stated reason. The slice is reproducible by a one-line predicate on the frozen corpus file. Results. computer-generated-content is the largest single inferred cause (4,704 records, 34.8%), ahead of compromised-peer-review (22.0%), plagiarism (18.6%), and editorial process (13.7%). The dominance holds on the title-level subset (41.8%) and the broad-CS-metadata tier (32.7%). 84% of these retractions are of papers published 2021–2023. Conclusions. Across both the title-level population (3,243) and the inclusive AI/ML population (13,502), computer-generated content is the largest project-inferred cause in the visible retraction record. The study is a characterization of public notices and metadata, not a misconduct census or a base-rate estimate. Clear, machine-readable retraction notices that name generation as the cause are the substrate for further research-integrity work.