About 90 Papers Retracted at Once, Most of Them Machine Learning

A sweep of Crossref's retraction records for the 24 hours to 27 August returns 101 notices. Roughly 90 of them come from a single publisher, IOP Publishing, across three conference series: Journal of Physics: Conference Series, IOP Conference Series: Materials Science and Engineering, and IOP Conference Series: Earth and Environmental Science. The remainder are scattered across MDPI, Springer, Frontiers, Wiley and one from Science Advances.

The subject concentration

Of the 101 titles, approximately 60% reference machine learning, neural networks, deep learning or artificial intelligence. That is not a coincidence of topic popularity. Conference proceedings series with high acceptance volume and light peer review have been the primary destination for paper-mill output for several years, and applied-machine-learning titles are the easiest to mass-produce: a standard architecture, a public dataset, a table of metrics, and a domain word swapped in.

What the common framing gets wrong

Each retraction is published as its own notice, with its own DOI, attached to its own paper. Read individually — which is how a researcher, a citation manager or an automated literature tool encounters them — every one of these reads as an ordinary editorial correction of a single flawed paper. Nothing in any individual notice states that 89 others were issued in the same window by the same publisher. The scale only becomes visible if you query the registry across a time range, which almost nobody does. The result is that a mass integrity action is delivered through a channel that structurally conceals that it is a mass action.

Why this is the wrong problem to under-report

Retracted machine-learning papers do not stop being cited. They sit in reference lists, in survey papers, and increasingly in the training corpora of the models being used to write the next round of literature reviews. A retraction notice that reaches only readers who open the specific paper it belongs to is a correction that does not propagate. When the retracted work is itself about machine learning methods, the failure to propagate feeds directly back into what models will assert about those methods.

What would fix the visibility

Nothing about the notices is deficient in itself — they are registered, machine-readable and correctly typed in Crossref, which is why this sweep is countable at all. What is missing is a summary act: a single publisher statement giving the number, the series affected, the investigation that produced it and the criteria applied. Without that, the only way to know that around 90 papers were withdrawn at once is to run the query, and the burden of noticing sits with the reader rather than the publisher.