AI content detection in scholarly publishing refers to the screening technology and editorial checks publishers use to flag manuscript text, images, or data that may have been produced or altered by generative AI without proper disclosure. Editors now combine AI-detection tools inside the STM Integrity Hub with disclosure requirements set by COPE and ICMJE, because no detection tool is accurate enough to make a final decision on its own. Before 2026 disclosure rules tighten further, editors need a documented process for flagging, verifying, and recording suspected undisclosed AI use — not just a tool that produces a score.
Why AI Detection Matters More in 2026
Generative AI text tools are now good enough that undisclosed use is often invisible to a reader on a single pass, which is exactly why screening has shifted from a manual editor’s instinct to a formal, tool-assisted step. Publishers report a steady rise in manuscripts showing templated phrasing, fabricated citations, or oddly generic discussion sections — patterns closely associated with paper mill operations that have adopted generative AI to scale their output.
The STM Association’s Integrity Hub, a shared screening system used by roughly 40 publishers, now processes more than 125,000 submissions a month across seven connected manuscript systems, intercepting around 1,000 suspected paper mill submissions monthly. That volume alone signals why individual editors can no longer catch every case through manual review: the problem has scaled past what one editorial desk can screen unaided.
At the same time, publisher policy has hardened. Major houses have converged on a shared position: AI cannot be listed as an author, and any substantive AI use must be disclosed in a specific manuscript location, whether that’s the methods section, the acknowledgements, or a standalone declaration. Editors who don’t have a documented, repeatable process for applying these rules are exposed on two fronts — missing genuine integrity problems, and inconsistently enforcing disclosure across authors and issues.
COPE, ICMJE and STM Integrity Hub Requirements
Three frameworks now sit behind almost every journal’s AI policy, and editors are expected to know how they differ.
COPE’s position on authorship and disclosure
COPE (Committee on Publication Ethics) is an international forum that publishes position statements on publication ethics. Its February 2023 position statement holds that AI tools cannot meet the requirements of authorship because they cannot take responsibility for the submitted work, hold conflicts of interest, or manage copyright and licence agreements.
Where authors do use AI — in writing, in producing images or graphical elements, or in collecting and analysing data — COPE requires transparent disclosure of the tool and how it was used, typically in the materials and methods section. Authors remain fully responsible for every part of the manuscript, including any AI-produced content, and are liable for any resulting breach of publication ethics.
ICMJE’s disclosure requirement
DefinitionICMJE (International Committee of Medical Journal Editors) is a group of biomedical-journal editors whose recommendations require AI use to be disclosed in both the cover letter and the manuscript, and state that AI tools cannot bear responsibility for a work’s accuracy, integrity, or originality.
ICMJE reaches the same conclusion as COPE from a slightly different angle, aimed specifically at biomedical journals. It requires disclosure in two places: the cover letter submitted with the manuscript, and the manuscript itself, in whichever section — methods or acknowledgements — best reflects how the tool was used.
What the STM Integrity Hub actually screens for
STM Integrity Hub is a shared, cloud-based screening system run by STM Solutions that lets scholarly publishers pool integrity signals — including papermill-similarity scoring and AI-generated text probability — to flag manuscripts for editorial review before publication.
It isn’t a single algorithm; it layers several purpose-built checks together at the point of submission, including duplicate and simultaneous-submission detection, papermill-similarity scoring through Clear Skies’ Papermill Alarm, and AI-assisted text analysis contributed by Springer Nature, tuned to catch templated language and other patterns typical of mass-produced papermill manuscripts. Every check produces a risk signal for an editor or integrity team to investigate — the Hub surfaces evidence; it does not issue an automated rejection.
| Framework | What It Covers | Where Disclosure Belongs |
|---|---|---|
| COPE | Authorship eligibility and disclosure obligations for text, images, and data | Materials and methods, or an equivalent section |
| ICMJE | Biomedical-journal authorship rules and submission-stage disclosure | Cover letter and the manuscript itself |
| STM Integrity Hub | Submission-stage screening for papermill and AI-generated content signals | Not a disclosure rule — a detection layer that feeds editorial review |
How AI-Content Screening Tools Actually Work
Most AI-content detection tools used in scholarly publishing work by scoring probability, not by proving intent. A manuscript is divided into sections, each section is checked for internal consistency, and a probability score reflects how likely that section is to be AI-generated. A higher score routes the manuscript to a human reviewer rather than triggering an automatic flag on the published record.
Springer Nature’s own detection tool, built with its Slimmer AI Science division and contributed to the STM Integrity Hub, illustrates this general approach. This design exists because every current detection method carries a real error rate in both directions. A tool can miss AI-generated text that has been lightly edited by a human, and it can also flag legitimate writing from a non-native English speaker or a heavily formulaic methods section as suspicious. That’s why the STM Integrity Hub’s own documentation is explicit that outputs are risk signals for an editor to investigate, not decisions in themselves.
A practical screening step for scholarly editors typically includes:
- Submission-stage scoring — text and image screening runs automatically when a manuscript enters the editorial system, before it reaches a handling editor.
- Threshold-based triage — low scores proceed normally; moderate and high scores are routed to a named reviewer or integrity contact rather than auto-rejected.
- Cross-checking against known patterns — papermill-similarity tools compare a submission against a database of confirmed fraudulent output, separate from generic AI-text probability scoring.
- A documented human decision — every flagged manuscript gets a recorded outcome: cleared, queried with the author, or escalated for a formal integrity investigation.
Manuscript preparation tools that clean up structure, references, and formatting before this stage — such as Siliconchips’ own EDGAR editorial automation tool — support this too: a manuscript with consistent metadata and verified citations is easier for screening tools and editors to assess accurately, since structural noise isn’t competing with genuine integrity signals.
Building Disclosure into the Editorial Process
Detection technology only closes half the gap. The other half is making disclosure a routine, unavoidable step rather than something authors remember only if asked directly. Journals that handle this well typically build in four checkpoints.
At submission, the system should require a direct declaration — not an optional field — asking whether generative AI was used in drafting, in image or graphic production, or in data analysis, and if so, which tool and for what purpose.
At the copyediting stage, editors should confirm the declared AI use matches what the manuscript actually shows and check that the disclosure statement sits in the correct section for that journal’s house style. This is also where trained editorial staff are best placed to spot fabricated or malformed citations that automated tools miss, since verifying whether a reference actually exists still depends on informed human review; Siliconchips’ copyediting services build this check into the standard reference-validation pass rather than treating it as a separate integrity task.
At peer review, reviewers should be told explicitly whether AI screening has already run and what it found, so they aren’t duplicating detection work the system already performed, and can instead focus on judging the substance of the research.
At the proof stage, the disclosure statement itself needs the same accuracy check as any other part of the manuscript — confirming it’s present, correctly worded, and consistent with what was declared at submission. Broader editorial support functions, including editorial support services, are well placed to own this final consistency check because they already sit across the full manuscript lifecycle rather than a single stage.
Where Human Judgement Still Has to Lead
No detection tool available in 2026 can reliably distinguish a well-disclosed, policy-compliant use of AI from an attempt to conceal it, because the underlying text can look identical. That distinction depends on context a machine doesn’t have: whether the author declared the tool, and whether the declared use matches journal policy. Three judgement calls specifically cannot be delegated to a screening score:
- Deciding whether a flagged section reflects genuine misconduct or a false positive — this requires reading the manuscript, not just the score.
- Weighing intent — an author who disclosed AI use in the wrong section made an editorial-formatting error; an author who didn’t disclose it at all raises a different, more serious question.
- Deciding what happens next — a query to the author, a correction, or a formal investigation is an editorial and ethical decision, and STM Integrity Hub’s own framing treats its output strictly as a signal for that decision, not a substitute for it.
The practical implication for editorial teams is that detection technology should reduce the volume of manuscripts needing close manual attention, not remove manual attention from the decision entirely.
Quick Glossary
- COPE (Committee on Publication Ethics): An international forum that publishes position statements and guidance on publication ethics, including its 2023 statement that AI tools cannot be listed as authors and must be disclosed when used.
- ICMJE (International Committee of Medical Journal Editors): A group of biomedical-journal editors whose recommendations require AI use to be disclosed in both the cover letter and the manuscript, and state that AI tools cannot bear responsibility for a work’s accuracy.
- STM Integrity Hub: A shared, cloud-based screening system run by STM Solutions that lets scholarly publishers pool integrity signals — including papermill-similarity scoring and AI-generated text probability — to flag manuscripts for editorial review before publication.
Frequently Asked Questions About AI Content Detection in Scholarly Publishing
Can AI content detection tools prove a manuscript was written by AI?
No. Detection tools produce a probability score based on textual patterns, not proof. A high score means a manuscript needs human review, not that misconduct has been confirmed. Editors are expected to verify any flagged section before taking action.
Do all publishers use the STM Integrity Hub?
No. Around 40 publishers currently connect to the Hub through seven integrated manuscript systems. Many other journals rely on their own in-house screening tools, third-party services, or manual editorial checks instead.
What happens if an author doesn’t disclose AI use and it’s later discovered?
Under COPE’s position, authors remain fully responsible for their entire manuscript, including AI-produced content, and are liable for any breach of publication ethics that results. Consequences vary by journal and severity, ranging from a required correction to a formal investigation.
Is AI-assisted copyediting the same as AI-generated content?
No, and most policies treat them differently. Using AI to check grammar or spelling is generally treated as a lower-risk editorial aid, while using AI to draft substantive text, analysis, or claims triggers the disclosure requirements set out by COPE and ICMJE.
Can editors rely on AI detection scores instead of manual review?
No. The STM Integrity Hub’s own documentation describes its output as a risk signal for an editor or integrity team to investigate, not an automated decision. Every major framework covered here keeps the final judgement with a human editor.
AI content detection in scholarly publishing works best as a triage layer, not a verdict. COPE and ICMJE set the disclosure rules; the STM Integrity Hub and similar screening technology surface the manuscripts that need a closer look; and trained editorial judgement decides what happens next. As disclosure expectations tighten through 2026, journals that build these checkpoints into a consistent, documented process — rather than relying on ad hoc editor instinct — will be better placed to catch genuine integrity problems without penalising authors who disclosed their AI use correctly.
See how EDGAR supports integrity-safe editorial operations
