Why did AI submissions surge after ChatGPT launched?

Organization Science's AI Task Force (Lamar Pierce, Claudine Gartenberg, Alex Murray and Sharique Hasan) reports that submission volume at the journal "has risen by 42% since the launch of ChatGPT in November 2022," compared with the prior two-year window — more than double the 20% bump the journal saw during the COVID-19 pandemic. The increase, they write, "is almost entirely due to manuscripts with substantial AI-generated text": submissions scoring below 15% on the Pangram AI-detection tool have declined since late 2022, while the 15-30%, 30-70% and 70%+ AI-score bands have grown enough to fill the gap and drive the aggregate rise. By early 2026, "the majority of manuscripts submitted to Organization Science contain detectable AI-generated writing," with the 70%+ band growing fastest. Average abstract readability (Flesch Reading Ease) fell steadily after the ChatGPT launch and by January 2026 sat 1.28 standard deviations below its January 2021 level. Among manuscripts scoring 70%+ AI, "nearly 70% are desk-rejected" without review. The task force frames this as the field being pushed "to produce more research rather than better research," worse on both writing quality and editorial outcomes.

Their mechanism for the readability drop is counterintuitive: AI writing is "less hedging, less passive, and more specific," which reads as polish, but it also carries "longer words, more complex sentence structures, more jargon, and more nominalizations," producing text they call "superficially clear but substantively impenetrable." They demonstrate this directly: Gartenberg stripped a published Pierce paper of its abstract and had ChatGPT 5.4 rewrite one for a "top" journal; the rewrite "makes sense, is grammatically correct, and yet is a slog to get through and much harder to grasp" than the plain original. Extending that logic to full manuscripts, they add a second mechanism for the desk-reject rate: heavy AI use signals that "authors substantially delegate writing, and the thinking that goes along with it, to the models," so the submission arrives without the clarifying work drafting normally forces — "the aha moments that come from the writing process are now gone."

This sits alongside Does LLM writing assistance change how scientists publish?, which measures a similar more-but-worse pattern across 2.1M preprints with a different detector; both find output rising while a quality signal that used to track effort — complexity there, readability here — stops meaning what it meant. This report adds an editorial-outcome measure, desk-rejection, that a preprint-level detector cannot reach, and ties the finding to one journal's actual gatekeeping decisions. It also complicates How widely do peer reviewers actually use AI tools?: Frontiers surveys reviewers' self-reported AI use and sees "untapped potential" for rigor, while this report measures what AI-heavy submissions do to the system those reviewers work in — adding volume and lowering readability, not raising rigor. It is also a milder instance of the same more-output-at-a-cost shape as Does AI help individual scientists while narrowing scientific focus?, where the cost is narrowed topic coverage rather than readability and desk-rejection.

The report does not establish how much of the readability drop or desk-reject rate is caused by AI writing itself rather than weaker underlying research arriving by the same route; the Pangram scores are measured on abstracts and only validated against, not measured across, full manuscripts, and the single worked rewrite is a demonstration, not a sampled comparison. It is one journal's submission pool over roughly a decade, not a field-wide or discipline-wide census, and the authors are self-described task-force members who say they are "in awe of the technology," a stance that does not bias the counts but does shape which findings they chose to report. The defensible reading is narrower than "AI makes research worse" in general: at Organization Science specifically, the submissions correlated with higher AI-detection scores are, on these two measures, harder to read and far more likely to be rejected without review, and reviewer and editor labor is absorbing a volume increase overwhelmingly composed of those submissions.