The AI Letter Flood: What the Surge in AI-Generated Correspondence Means for Medical Authors

For most of academic publishing history, letters to the editor have been a minor corner of the scientific record: short, rapid-turnaround commentary from researchers who had something to add to a published paper and wanted to say it in print. The process was lightweight. Editorial review was faster and less rigorous than for original research. Most letter writers published a handful over a career.

That corner is now a problem. A preprint analysis covering PubMed-indexed publications from 2023 to 2025 found that roughly 8,000 authors moved from the bottom to the top of the productivity distribution for letters, going from publishing few or none to publishing many in a very short period. This group made up only 3% of all active letter authors during that window, but accounted for 22% of all letters published, nearly 23,000 in total. Their work appeared in 1,930 journals, including 175 letters in The Lancet and 122 in The New England Journal of Medicine. One physician, based in Qatar, published more than 80 letters in 2025 after publishing none in 2024.

Researchers and editors who have looked at this pattern say it is almost certainly linked to large language models. The analysis was covered by Science in late 2025, and editorial commentary in the journal PRiMER described a specific incident from summer 2025 in which the editorial team received and identified an AI-generated letter. The problem is not hypothetical anymore. It is in the indexes.

Why the Correspondence Section Is Especially Vulnerable

Letters to the editor are handled differently from original research at almost every journal. Word limits are tight, usually 500 to 750 words. Turnaround is fast, sometimes days rather than months. Editorial scrutiny is lighter. Statistical review does not happen. Reference verification is typically informal. All of this makes the section efficient for its intended purpose, which is rapid scholarly exchange. It also makes it structurally easier to exploit.

The workflow for AI-assisted letter generation is not difficult to imagine. A language model can scan a recently published abstract, identify points of departure, generate a nominally critical or supportive commentary, and format it with a reference list in a matter of minutes. Because letters rarely exceed a page, the surface area for hallucinated content is small enough that it can look plausible without careful checking. A fake reference or two tends to get caught only if someone tries to locate the cited paper, which busy editorial offices often do not have time to do.

The incentive structure compounds this. As one researcher quoted in the Science coverage put it, it is easier to get a letter published in NEJM than a paper, and some institutions count them the same in promotion dossiers. For researchers operating under strong publish-or-perish pressure, particularly in systems where any indexed publication carries value regardless of type, a rapid-fire letter operation looks rational even if the underlying output is thin. That calculation does not require conscious fraud. It requires only a willingness to use a tool without thinking carefully about what the tool produces.

The Intensive Care Medicine Case

The most documented individual example to date appeared in Intensive Care Medicine, a Springer Nature journal. A letter published in December 2024 explored how AI might assist clinicians in monitoring hemodynamic status in intensive care patients, a genuinely relevant topic. The 750-word piece carried 15 references.

When Retraction Watch examined those references, only five could be verified as real publications, and even those had discrepancies in author order, publication year, or title. The remaining ten references could not be located in the cited journals or in any journal at all. One of the references was entirely fabricated, citing a paper purportedly published in Intensive Care Medicine itself, a journal the letter was addressing, as its source. That paper does not exist.

The authors stated, after the problems were identified, that the fabricated references arose from using a generative AI tool to convert PubMed IDs into a structured reference list, a task language models handle poorly because they tend to confabulate bibliographic details when the input is incomplete. The journal added an editor's note on November 4 expressing concern about the references. The letter was retracted on November 29. The sequence from publication to retraction took roughly a year.

This case is instructive not because it is particularly egregious but because it is so routine. The letter addressed a legitimate topic. The authors were real. The error, by their own account, was treating an AI output as reliable without manual verification. That gap between what AI tools produce and what authors verify before submission is where most of the current correspondence integrity problem lives.

What the Prolific Debutante Pattern Reveals About Authorship Norms

The phrase "prolific debutante" was coined by researchers analyzing the letter output data to describe authors who appear suddenly at the top of the productivity distribution without a corresponding career arc. Established letter writers tend to publish slowly and selectively, responding to papers in their primary field. The debutante pattern is different: many letters, across multiple journals and topics, published in a compressed period, often with no prior track record in original research.

Not every new, highly productive letter author is gaming the system. Some researchers genuinely accelerate their correspondence output. But the statistical pattern across 8,000 authors moving in the same direction, in the same three-year period immediately following the wide release of large language models, is hard to explain otherwise. The 22% share of total letters is particularly striking when you consider that these authors represented only 3% of the active author pool.

The topic clustering matters too. Researchers who have looked at the letters from these accounts note that they tend to cover newly published papers broadly rather than specializing in a single niche, which is the opposite of what you would expect from a researcher with genuine expertise. The commentary is often technically coherent but thin on specificity, and references, where they can be verified, are sometimes loosely connected to the actual claim being made.

This pattern also raises a different kind of integrity question. When AI-generated commentary crowds the correspondence section of a journal, it competes with and potentially displaces genuine expert response. A researcher who has real criticism of a published trial, and who takes the time to write a substantive letter, may find the section cluttered with generic-sounding commentary that received the same editorial attention without the same substance. That displacement has a cost for the scientific discourse journals are supposed to host.

How Publishers Are Responding

The publishing industry's response to AI-generated content has accelerated across 2025 and into 2026, though most of the tools being deployed were designed for original research rather than short correspondence. Springer Nature has built or deployed nearly 60 AI-based tools across its editorial workflow, covering manuscript screening, peer reviewer matching, editor evaluation support, and research integrity checking. The company reported in early 2026 that these tools processed more than 1.5 million research papers in 2025, flagging around 25,000 for potential issues including fabricated text, invented references, and image manipulation.

Two of these tools are particularly relevant to the text fabrication problem. Geppetto, developed in partnership with Slimmer AI Science, detects AI-generated content by analyzing textual consistency across sections and scoring the probability that text in each section originated from a language model. SnappShot, the image counterpart, analyzes gel and blot images for duplication. Springer Nature has also donated an earlier version of its text detection tool to the STM Integrity Hub, a cross-publisher consortium that lets member publishers access shared integrity tools for screening.

The BMJ and NEJM both require disclosure of AI use in correspondence as they do in original research. NEJM allows AI assistance for writing but requires explicit acknowledgment in the cover letter and the letter itself. BMJ's AI disclosure requirements extend to the submission workflow, where authors must confirm whether AI tools were used at any stage. Despite these policies, estimates suggest that only around 0.1% of post-2023 publications explicitly disclose AI use, even in journals with formal requirements. The disclosure gap is wide.

The more direct response to the volume pattern has come at the editorial level rather than the policy level. Some journals have quietly tightened their correspondence review process, adding reference verification steps or requiring authors to confirm each cited paper was checked. Others have introduced limits on the number of letters a single author can submit within a rolling period. These measures are reasonable, but they are not published as formal policy at most journals, which makes them invisible to authors preparing submissions.

The Limits of AI Detection in Short-Form Content

The AI detection tools currently available, including Geppetto and commercial services used by multiple publishers, are trained primarily on full-length manuscripts. Short correspondence of 500 to 750 words does not give these tools much text to analyze, and the statistical signals that indicate AI generation at document scale are harder to read in a brief piece. A letter that is 60% AI-generated and 40% manually edited may not score high on current detection metrics even if the problematic content is concentrated in the reference list or the conclusion.

Reference verification remains the most reliable detection mechanism for the specific problem the Intensive Care Medicine case illustrated. Checking whether cited papers exist, whether they say what the letter claims, and whether they have the bibliographic details listed is slow manual work. No tool currently automates this completely at scale, though Crossref's reference verification API, Retraction Watch's database, and scite.ai allow automated checking of whether a DOI resolves to a real paper and whether the paper has been retracted. Publishers who integrate these checks at submission catch fabricated references earlier. Many do not.

There is also the question of what detection should accomplish. Identifying that a letter contains AI-generated text does not, by itself, determine whether the commentary is scientifically worthless. Some AI-assisted letters are accurate and genuinely respond to the paper they discuss. The problem is not AI assistance per se. The problem is undisclosed AI assistance combined with insufficient human verification of what the tool produced, especially in the references. That distinction matters for how journals frame their responses, and for how authors should frame their own practices.

What This Means for Authors Whose Research Gets Commented On

If you have published original research in a major medical journal in the past two years, there is a meaningful chance that some of the letters responding to your paper were generated with substantial AI assistance. Some of those letters will have contained inaccurate characterizations of your methods, cited papers that do not exist, or made claims about your findings that a careful human reader would not make. You may have received invitations from the journal to respond to these letters, or editors may have decided to publish them alongside a response you did not know was coming.

This is worth knowing because it changes how you should read and respond to correspondence about your work. A letter that seems to misrepresent what you found is worth checking carefully before you invest time in a rebuttal, because the misrepresentation may not reflect a genuine misreading but rather an AI output that paraphrased your abstract incorrectly. Similarly, a letter with an extensive reference list is worth spot-checking before you assume its literature review reflects the actual state of the field.

Authors who write letters have a parallel obligation. If you use AI assistance to help draft a response to a published paper, you should verify every reference you include by checking that the paper exists, that it says what you claim, and that it has not been retracted. You should disclose the AI use in line with the target journal's policy. And you should review the letter yourself for claims that the AI may have inflated, softened, or invented. Short as the format is, it goes into the permanent record.

The Institutional Incentive Problem Behind the Volume

The prolific debutante pattern cannot be explained by individual bad actors alone. Individual behavior at this scale is driven by systems, and the system relevant here is the publication-counting framework used for academic promotion in many institutions, particularly in countries where indexed publication counts carry heavy weight in hiring and career advancement decisions. When a letter in The Lancet and a letter in a regional specialty journal both count as publications, and when neither requires the same depth of contribution as an original research paper, the conditions for volume-oriented behavior exist independent of AI.

AI has lowered the marginal cost of producing each unit in that volume nearly to zero. A researcher who previously might have written two or three letters per year based on genuine engagement with the literature can now produce dozens with the same time investment, if the goal is output rather than contribution. That is not a technology problem. It is an incentive problem that the technology has made dramatically cheaper to exploit.

Journals are unlikely to solve this alone. They can tighten review, add reference checks, and impose per-author submission limits, but as long as the underlying counting framework assigns value to volume over quality, the pressure will find new channels. The more durable fix is at the institutional level, where promotion committees need to distinguish between first-authored original research, collaborative research, and correspondence, and between correspondence that advances genuine scientific exchange and correspondence that fills a productivity metric. That conversation is happening slowly in some academic communities and not at all in others.

Practical Steps for Authors Writing or Responding to Letters

If you write letters to journals, the current environment puts more responsibility on you than it did three years ago. The checklist below covers what a careful author should do before submission.

If you are responding to a letter commenting on your own published research, the same verification logic applies in reverse. Before writing a point-by-point rebuttal, confirm that the references in the incoming letter are real. If they are not, that is information the journal editor needs to know. Post a query directly to the editorial office rather than addressing a phantom citation in your response, which only extends its life in the record.

Where the Correspondence Section Goes From Here

Medical journals have been debating whether the correspondence section remains worth the editorial overhead for some years before this problem arrived. Many journals quietly stopped accepting unsolicited letters during the pandemic and did not restart. Others moved to online-only commentary with lighter editorial handling. The AI correspondence surge is likely to accelerate those changes at journals that have not already made them.

The journals that retain a correspondence section in print will probably tighten it further. Expect reference verification to become standard rather than selective, per-author submission limits to be formalized in author guidelines, and AI disclosure to become a required field in correspondence submission forms, not just a policy statement on the author instructions page. Whether those measures reduce the volume of AI-generated correspondence is uncertain. They will raise the cost of submitting one and may deter the least careful cases.

The deeper question is what the correspondence section is for. At its best it is a rapid, expert response loop that improves the scientific record by catching errors, raising alternative interpretations, and connecting a published paper to the broader literature. That function is genuinely valuable. Whether the AI correspondence wave has degraded it significantly depends on how much of the 22% figure represents genuine, if AI-assisted, scholarly commentary, and how much represents volume-maximization with no real scholarly intent. Nobody has counted that yet, and it may be difficult to count, because the outputs can look similar on the surface.

For now, the practical implication for working researchers is clear. If you value the correspondence sections of the journals you read and publish in, the response is to hold your own letters to a higher standard, not a lower one. Write them because you have something specific to say about a specific paper, verify everything you cite, and disclose any AI help you use. That is how the format stays worth reading.

Further Reading