The paper got cheap
I submitted the manuscript at 7:32 on a Saturday morning, through my phone's hotspot, because my home internet provider cannot reach the publisher's login page. It was a benchmark of five forecasting models on fourteen years of monthly cardiovascular deaths in São Paulo: classical statistics, gradient boosting and a foundation model, with a bootstrap interval around every comparison. Four authors had signed off. The rejection arrived the same day, before any reviewer saw it.
The editor was unusually frank. Submissions had doubled since the summer, on top of a 40% rise the year before, and the journal can only publish 100 to 150 papers a year. The editor could see that the paper had been written, and largely built, with AI agents, and judged the contribution too thin "in view of what can be achieved with agentic AI."
They were right on both counts, and I can show it. That evening I ran a script over the manuscript. Seven thousand words, twenty-five uses of "rather than". Eight section titles that stated a conclusion, like small press releases. A Methods section that told the story of how we had debugged our own pipeline, which is what an agent does when it turns its session log into prose. Fifteen reminders of how carefully we had fixed our criteria in advance. The word "survive" six times.
The science had the same problem as the prose. Five off-the-shelf models on one public series is an afternoon of work for a coding agent. We did it carefully, with intervals and paired tests. Careful is now the floor.
I understand why this stings. The paper has been the unit of a scientific career for a long time, and for a PhD student the CV is a list of them. Peer review was already strained, with editors chasing reviewers for months. Now the supply of competent-looking manuscripts is growing much faster than the supply of people able to read them.
Here is how I think about the agent now. It is a coauthor who is very fast, never tired, has read everything, and has never met a patient. It has never sat with the team at a state health department that decides how many coronary care beds to staff in July. It does not know that registries in parts of the North and Northeast still miss deaths that São Paulo records, or that the share of ill-defined causes moves with the quality of the local system. It will write a beautiful Methods section for a question nobody needed answered.
So the benchmark got cheap, and so did the figures, the prose, the literature summary and even statistics done properly. What stays scarce is the question, data that cost someone real work to collect or link, design decisions that only domain knowledge can defend, and the voice of whoever will use the result.
From one angle this looks bleak. Editors are learning to spot agent fingerprints in thirty seconds, and some good work will be thrown out with the bad. Small labs in Brazil already fight for attention in journals edited in Europe and the United States, and a new reason for a desk reject does not help them.
Then look at what the agent did for us the same night. Our own Discussion had said that a foundation model should matter most where series are short and many: municipalities, age groups, newly defined causes. We had never tested it, because building and checking hundreds of series by hand was beyond a small team. It no longer is. The second version covers the 27 Brazilian states, four groups of cardiovascular causes, sex and age, close to 650 series, with population rates and calibrated intervals as the main outcome. The download of fourteen years of death certificates started a few hours after the rejection. A small team can now ask the question that used to need a consortium.
Two kinds of scientists should worry. The first uses agents to produce the same papers faster, which is the mistake we made; editors and readers can tell. The second refuses to use them, spends months on what a colleague does over a weekend, and still competes for the same 150 slots.
The agent can write the paper now. Bring it a question worth the trouble.