AI promises to free editors from repetitive work. But what if those tasks also teach editors how to judge?
Imagine two junior editors beginning their careers at university presses.
One reads incoming manuscripts, checks references, identifies weaknesses in arguments, and discusses difficult submissions with senior colleagues. Much of the work is repetitive, occasionally frustrating, and invisible to readers.
The other joins a press where AI screens submissions, prepares preliminary assessments, highlights suspect citations, and recommends manuscripts for further consideration. This editor spends less time on initial checks and more time reviewing material already processed by intelligent systems.
Five years later, both may be productive. The second may even handle far more submissions. But which has developed stronger editorial judgment?
There is no automatic answer. AI might help the second editor learn faster by exposing weaknesses and alternative interpretations. Yet the comparison reveals an uncomfortable possibility: some of the work publishers are eager to automate is also the work through which editors develop expertise.
As AI moves deeper into publishing, the question is no longer whether it can assist editorial decisions. It is whether publishers can make work faster without making their editors less capable of judging it independently.
What the Efficiency Argument Leaves Out
The appeal of automation is understandable. Scholarly publishers face large submission volumes, limited resources, and pressure to reduce delays without compromising quality.
In March 2026, Springer Nature reported that more than 1.5 million papers benefited from nearly 60 AI tools in 2025. Its Editor Evaluation tool supported assessment of nearly half a million manuscripts, while other tools flagged 25,000 papers with possible problems involving images, references, or fabricated text.
These systems can relieve genuine pressure. Finding missing documents, extracting metadata, checking formatting, and surfacing potential integrity concerns are useful forms of assistance. Few editors would argue for spending more hours on avoidable administration.
The familiar promise is that automating routine publishing tasks gives editors more time for intellectual judgment. That is plausible, but incomplete. It assumes judgment remains intact when the work surrounding it disappears.
Editorial expertise is not a resource stored safely away until someone has time to use it. It develops through repeated encounters with real manuscripts, including difficult and disappointing ones.
An editor learns to recognize a weak argument partly by reading dozens of weak arguments. A copyeditor learns to preserve an author’s meaning by wrestling with awkward sentences rather than merely accepting polished alternatives. Checking references can reveal that a seemingly persuasive claim rests on irrelevant or nonexistent evidence.
Of course, not every routine activity teaches something important. Confirming that a file is named correctly is different from testing whether a citation supports an argument. The challenge is distinguishing administrative repetition, which can usually be removed, from formative repetition, which builds professional judgment.
Publishers that automate both without distinction may win immediate efficiency while weakening the expertise on which future decisions depend.
When Editors Start With Someone Else’s Assessment
AI changes more than the volume of work. It can change the order in which thinking happens.
Traditionally, an editor encountering a manuscript begins with questions: What is the argument? Is the evidence persuasive? What has the author overlooked? Does the work offer something new?
In an AI-assisted workflow, answers may arrive first. The editor receives a summary, a list of weaknesses, a suitability score, or a recommendation. The task subtly shifts from constructing an independent interpretation to evaluating one already supplied.
That matters because algorithmic systems can shape which editorial questions get asked before a human makes the decision. An editor can reject the machine’s conclusion, but its initial framing may still influence what receives attention.
There is some relevant evidence from knowledge work more broadly. A 2025 study by Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers about 936 examples of AI-assisted tasks. Greater confidence in AI was associated with less self-reported critical thinking, while greater confidence in one’s own abilities was associated with more.
This was not an experiment showing that AI causes editorial deskilling. The researchers studied self-reported experiences across different professions. Still, they identified a shift worth examining: from gathering information and solving problems toward checking and integrating AI-generated responses.
For an experienced editor, scrutinizing a machine-generated assessment can be demanding intellectual work. For a novice who has rarely made an unaided assessment, it may become the only kind of editorial work they know.
Three Risks That May Accumulate Over Time
The apprenticeship gap
Editorial judgment is partly learned through apprenticeship. Junior editors read imperfect submissions, defend their recommendations, receive corrections, and gradually learn which problems matter most.
If preliminary screening and substantive checks are performed primarily by AI, some editors may receive fewer opportunities to practice these skills. They can become skilled at interpreting recommendations without gaining equal experience producing their own.
This outcome has not been conclusively demonstrated in publishing. It is a risk publishers should test, not a decline they should assume has already occurred.
The key question is practical: could a newly promoted editor assess a difficult manuscript confidently if the AI system were unavailable? If not, that is a training problem even when day-to-day productivity looks excellent.
Agreement mistaken for evaluation
A polished AI assessment can be persuasive. The editor reads it, finds the reasoning plausible, and agrees. The decision remains formally human, yet agreement alone tells us little about the depth of independent scrutiny.
An unconventional interdisciplinary paper, for instance, might look unsuitable under familiar categories while making a valuable contribution. An apparently strong argument might contain a conceptual flaw that automated checks fail to recognize.
Editors need enough independent knowledge to notice omissions, question criteria, and resist recommendations that look reasonable but are wrong. Without those abilities, a human sign-off can become a procedural safeguard rather than an intellectual one.
Verification without an owner
Publishing relies on overlapping checks, but multiple checkpoints do not guarantee that essential verification happens.
In 2025, Springer Nature retracted Mastering Machine Learning: From Basics to Advanced after the publisher could not verify 25 of its 46 references, according to Retraction Watch.
The case does not establish that AI caused the editorial failure. It demonstrates the consequences when serious problems survive a publication process.
Meanwhile, a Frontiers survey published in December 2025 found that 53% of peer reviewers reported using AI tools, often for summaries or review-writing assistance. Such support can be useful, provided confidentiality rules are respected and reviewers verify their own work.
The danger comes when an author, reviewer, editor, and automated system each assume somebody else has checked the evidence. Responsibility can become dispersed even while every stage appears to have been completed.
Human Judgment Is Not a Perfect Alternative
There is a danger of romanticizing editorial work before AI. Human editors make mistakes, reviewers disagree, and established conventions can unfairly privilege familiar institutions, methodologies, or writing styles.
AI may help correct some weaknesses. It can apply administrative criteria consistently, identify overlooked anomalies, and offer alternative interpretations that challenge an editor’s first impression. An editor who thoughtfully compares independent reasoning with AI feedback may become better at the job.
Nor is all AI-assisted learning inferior to traditional apprenticeship. An appropriate tool can explain why a citation fails to support a claim or reveal patterns an inexperienced editor would otherwise miss.
The objective is therefore not to insist on manual work for its own sake. It is to preserve a capacity for judgment that is independent enough to challenge AI and informed enough to benefit from it.
That requires more than declaring that humans remain in control. Publishers need workflows that make critical engagement possible, visible, and teachable.
Designing Workflows That Protect Editorial Expertise
Let editors form a view before seeing the recommendation
For consequential decisions, especially those involving originality, publication suitability, or rejection, an editor could record a brief preliminary assessment before opening the AI report.
This need not happen for every routine check. It can be reserved for complex cases, borderline submissions, training exercises, and occasional quality audits.
The sequence matters. When the editor’s assessment comes first, differences become opportunities to investigate rather than reasons to conform automatically.
Keep formative tasks in editorial training
Junior editors should continue to read complete manuscripts, verify selected references, perform substantive edits, and explain their conclusions to experienced colleagues.
Publishers could set aside periodic AI-free exercises or rotate staff through manual assessment. Then, after an independent evaluation, an AI tool could provide a second perspective.
This is not nostalgia for slow workflows. It is deliberate practice in skills that remain necessary when automation fails, oversimplifies, or encounters unfamiliar scholarship.
Separate information from recommendation
A system that identifies missing ethics statements has a different influence from one that assigns a manuscript a quality score or suggests rejection.
Publishers should map where AI merely supplies evidence and where it begins to frame the decision itself. The latter uses warrant greater scrutiny, including testing for errors and unintended bias.
This is especially important for regional and multilingual publishing, where an AI tool trained around dominant scholarly conventions may misread unfamiliar language use, disciplines, or research contexts.
Audit judgment, not just speed
Turnaround time and staff hours are easy to count. The strength of editorial reasoning is harder to measure, but not impossible to examine.
Publishers can compare unaided and AI-assisted assessments, review disagreements, and test whether editors detect deliberately introduced problems. They can also study when editors override recommendations and whether their explanations hold up under review.
A low override rate does not automatically prove overreliance; the system might be accurate. The useful question is whether editors can explain agreement as convincingly as disagreement.
Give verification clear ownership
Every AI-assisted step should specify who is responsible for checking its output. If a tool flags a reference, someone must decide whether the concern is valid. If reviewers use permitted AI assistance, they must still verify the substance of their reports.
Confidential submissions also require safeguards: publishers should not allow unpublished manuscripts or reviewer materials to enter tools without appropriate authorization and security protections.
Human accountability becomes meaningful only when the necessary work is actually assigned and performed.
The Test Publishers Should Apply
AI can improve publishing without diminishing its editors. It may reduce avoidable work, provide useful second opinions, and make room for deeper evaluation. But these benefits do not guarantee that editorial expertise will maintain itself.
The more revealing measure may emerge years later, when today’s junior editors become tomorrow’s senior decision-makers.
Will they know how to recognize an important contribution that the system overlooks? Can they identify a plausible but unsupported claim? Can they defend a rejection without repeating an automated assessment? Most importantly, can they explain when and why the machine is wrong?
If they cannot, publishers may have created remarkably efficient workflows while neglecting the profession those workflows are supposed to support.
The aim should not be to preserve every manual task. It should be to preserve the learning, independence, and accountability embedded in the tasks that matter.
Publishers will know they have embraced AI successfully not simply when editors can process more manuscripts, but when those editors remain capable of making better judgments, even without the machine.