Should our journal publish AI-drafted manuscripts?

Forget both truth and beauty, I want to know about opportunity costs Status: Rough conceptual model. This is a personal exploration of a live policy problem and definitely does not represent the opinion of the Alignment Journal itself. Given the context, I had best disclose my own AI usage in this article: transformative. Although the original model design was mine, it was made way better by iterative refinement and re-drafting by AI, and by no means would I have had time to write it purely by hand. At the Alignment Journal we have been discussing whether to accept AI-drafted manuscripts for review. This is a relatively high-leverage question, as detecting AI-drafted prose is (currently, against authors who are not trying to hide it) surprisingly feasible, making it a cheap (albeit imperfect) proxy signal for us to use in desk review to filter out low-quality papers.[1] The ideal policy would optimize for the overall quality of the journal’s output with regard to how that serves our readers. There are many components to that; quality, readability, professional and academic norms… I ignore most of those and bloody-mindedly focus on the economic effects of AI drafting. On one hand, AI drafting lowers the cost of producing a manuscript, which may increase the flow of good work into our readers’ inboxes. On the other hand, it might instead increase the flow of low-quality AI slop into those inboxes. That is, we worry about the impact AI slop spam, as do many others (Gartenberg et al. 2026). Insofar as the base costs of a high-quality paper exceed those of a slop manuscript, we might hope that permitting AI drafting provides a relatively larger benefit to slop authors than to substantive authors. So, if we can detect AI drafting, should we do so? And: should we use it to desk-reject papers? Here is a minimal economic model of that question, in the vein of Agrawal, Gans, and Goldfarb (2019), which models AI as a fall in the price of prediction. My assumptions are stylized and tractable rather than realistic or empirically calibrated. Where I need to choose some numbers, I have taken inspiration from the Alignment Journal, since that is closest to my heart. Here, I assume AI drafting reduces the cost of producing prose: the writing-up stage of science gets cheap while the substance stage, for now, does not (as per Kwa et al. 2025). That is a shaky assumption even now (for example, both my drafting and substantive work in writing this very post were AI-expedited) and it will get shakier as AI progresses. My goal here is not to persuade anyone of a particular optimal policy, but to lay out a concrete model of the first- and second-order economic effects of AI drafting so that we can debate these things better. Too often the discussion is conducted in purely moral terms (AI is “efficient” or “lazy”) or aesthetic ones (“AI prose is hard to read”); those lenses matter, but our duties in publishing surely also include the economic and strategic management of scientific knowledge production, and quantitative models of that seem under-supplied. tl;dr Under my (contrived but not crazy) assumptions: AI drafting preferentially incentivises production of poor papers, which crowd out good ones: the time it saves good authors turns into extra manuscripts that are never read. A ban enforced by an AI-drafting detector can induce three different regimes, depending on detector quality: At low true positive rate it changes nothing; at intermediate levels it deters only the good authors, who give up AI before slop authors; then, if the detector is good enough, it restores the no-AI outcome, forfeiting every hour AI would have saved. A flat submission fee lets everyone use AI while excluding spam, and beats the ban outright. The fee that does it is ballpark USD 2,500 to 7,500 per submission, which is outside the Overton window and moreover a business model indistinguishable from a predatory journal’s, so nobody, least of all me, is proposing it. A hybrid, where authors may pay to skip the detector, charges only AI drafters and publishes more good papers than the ban but fewer than the flat fee, because slop authors who draft by hand still clog the queue. The fee that does it is USD 4,000 to 12,000 per AI-drafted submission if the detector holds up against motivated rewriting.. Not modelled: false positives, reputational effects, harmful slop, competing journals, and AI that does the science as well as the writing. And, clearly, I remain silent on whether AI can do good writing or not. Putting that all together we can see that, regarded purely as an instrument for controlling slop rates, the detector ban is a clumsy approximation to a congestion charge, but it might be better than nothing and is more palateable than fees would be. Setup I assume there are two types of authors: high-quality authors (type mjx-mover { display: inline-block; text-align: left; } mjx-mover:not([limits="false"]) { padding-top: .1em; } mjx-mo