TMLR Editor Asks Authors About Their Own Papers; All 10 Desk-Rejected Submissions Rejected

On September 16, 2026, the machine learning journal TMLR published "Asking Authors About Their Own Papers" on its official blog. The author was Nihar B. Shah of Carnegie Mellon University, one of the journal's co-editors-in-chief. During his editor-in-chief rotation from August 14 to 28, 2026, he informally selected 10 submissions that were slated for desk rejection and, via OpenReview messages, invited their authors to "have a chat" instead of rejecting them outright. In the end, all 10 papers were still rejected.

Divergence quickly emerged: the author of one paper withdrew it directly, one author replied that they had no time for a call, and the remaining 8 were scheduled for meetings, one of whom did not show up on time. In the end, Shah met with the authors of 7 papers, whose identities included undergraduates, master's students, doctoral students, university faculty, and independent researchers; most were single-author papers, but not all. Shah's questions fell into two categories: basic questions about the problem setup, notation, and claimed results, and detailed questions about technical expressions, theoretical results, and experimental design choices.

Of the 7 meetings, only 1 author answered all questions; another 3 could explain the high-level ideas clearly but ran into difficulty as soon as technical details were probed; the worst 3 could not answer even basic questions, and all were single-author papers. Shah wrote that two of these authors appeared to have almost no substantive understanding of their own papers' content, and another could not find where several key results claimed in the abstract were actually given or supported in the main text. The one paper whose author answered everything correctly also did not pass: during his review, Shah found a major error in one of the main conclusions, which the author subsequently acknowledged; TMLR desk-rejected it but allowed resubmission after correcting the error or narrowing the conclusion. The other 9 were desk-rejected with no resubmission option.

Shah also recorded two things. First: two authors who could not answer basic questions at the meeting later sent written responses, and Shah fed these two emails to the AI text detection tool Pangram, which judged both as "100% AI." Second: in another meeting, an author tried to introduce the analysis method they had used, and in the course of explaining, inadvertently described a complete p-hacking workflow—that is, repeatedly adjusting the analysis approach until the data became "significant."

It should be noted that Shah himself does not reject AI. In the article, he admitted that to read 8 papers within two weeks, he himself used LLMs to assist his understanding, and even learned some concepts he had previously been unfamiliar with. What concerns him is not whether authors used AI, but whether authors can take responsibility for content bearing their names.

The experiment itself has clear limitations: the sample of 10 is very small, it was "informally" selected by Shah, and all came from papers that were already slated for desk rejection. What it can show is "what is present among submissions judged to be low quality," not the overall picture of TMLR submissions. Shah himself wrote that one purpose of the experiment was precisely to test whether the desk rejection process is reliable.