Authors are asked to declare every use of AI. What about the machines now reading manuscripts on the publisher’s side of the desk?
Comparative analysisEight major publishersFour transparency dimensionsAugust 2026
Every conversation about artificial intelligence in scholarly publishing seems to circle the same question: did the authors disclose their use of AI? It is a fair question. But it quietly assumes that the only AI in the room belongs to the author. It does not.
By 2026, nearly every major publisher screens submitted manuscripts with AI-assisted tools before a single peer reviewer is invited. Similarity checks, image-integrity analysis, reference verification, paper-mill signals, reviewer matching: this is legitimate, often necessary, integrity work. Yet a striking asymmetry has emerged. Authors are expected to disclose their AI use down to the tool, the version, and the purpose, while a publisher’s own editorial AI use typically remains visible only in a general policy page, and invisible at the level of the individual manuscript.
01The pattern
Institutional transparency, article-level opacity
This report compares eight major publishers on the transparency of their AI-assisted editorial screening: Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE, Frontiers, ACS, and IEEE. Across all of them, one pattern dominates. Publishers now openly acknowledge that they use AI-assisted screening at the corporate level. What they rarely disclose, at the point of submission and for a specific paper, is which tool was applied, how the manuscript’s data was handled, how heavily an algorithmic flag weighed on the editorial decision, and whether an author has any recourse designed for AI-caused error.
The aim here is not to single anyone out. Screening manuscripts for fabricated images, tortured phrasing, or paper-mill fingerprints is a defensible response to real threats. The argument is narrower and, we hope, constructive: transparency deserves to be treated as a three-way obligation shared by author, reviewer, and publisher, rather than a duty that runs in one direction only.
02How we assessed it
Four dimensions of transparency
Each publisher was evaluated against four dimensions, chosen to align with established principles of algorithmic governance: transparency, contestability, and meaningful human oversight.
D1 — Disclosure
Is AI-assisted screening declared?
Whether the publisher publicly states that AI-assisted tools are used to evaluate submitted manuscripts.
D2 — Human oversight
Is a human kept in the loop?
Whether the publisher assures that algorithmic output flags rather than decides, with a human expert retaining authority.
D3 — Data & third parties
Where does the manuscript go?
Which tools are used, whether in-house or third-party, and whether manuscript data is retained or used to train models.
D4 — AI-specific appeal
Can an author contest a flag?
Whether the author can learn they were flagged, see the output, and request human re-review of an AI-driven signal.
03The comparison
Where each publisher stands
Table 1 reads as a formal transparency ledger. Every publisher discloses AI use in some form and every publisher asserts human oversight; the divergence appears in data governance and, most uniformly, in the absence of any AI-specific route of appeal.
Table 1
Transparency matrix across four dimensions
| Publisher | D1 Disclosure | D2 Human oversight | D3 Data / third party | D4 AI-specific appeal | Maturity |
|---|---|---|---|---|---|
| Frontiers | Met | Met | Partial | Not met | High |
| Springer Nature | Met | Met | Partial | Not met | High |
| Elsevier | Met | Met | Partial | Not met | High–Mod. |
| Taylor & Francis | Met | Met | Partial | Not met | Mod.–High |
| Wiley | Met | Met | Partial | Not met | Moderate |
| ACS | Partial | Met | Not met | Not met | Low–Mod. |
| IEEE | Partial | Met | Not met | Not met | Low–Mod. |
| SAGE | Partial | Met | Not met | Not met | Low |
MetPartialNot metThe D4 column is uniformly “Not met,” making the absence of an AI-specific appeal mechanism the single most consistent finding of this analysis.
The publishers in brief
Frontiers is the most transparent case. Its in-house assistant, AIRA, screens every submission before peer review, running more than forty checks (language quality, image integrity, plagiarism, ethics compliance, paper-mill signals, conflicts of interest). Frontiers states plainly that AIRA flags but does not decide. What it does not fully publish are thresholds, false-positive rates, and a detailed route of redress for an honest author flagged in error.
Springer Nature no longer conceals the scale of its AI use: it reports that in 2025 more than 1.5 million papers were supported by roughly sixty AI tools across screening, editorial evaluation, and integrity checks, with named tools such as Geppetto (AI-generated text) and SnappShot (image manipulation), and states that each result is re-checked by a human expert. What remains opaque is which journal uses which tool, whether full text or derived features are processed, and each tool’s error rate.
Elsevier pairs a clear policy with the 2026 rollout of its Check Integrity tool to roughly two thousand journals, and defines a “private” AI tool as one that does not store, reuse, share, or train on input. Still, an author cannot see which tool touched their paper, what flags it produced, or how those flags bore on a desk decision.
Taylor & Francis is notably specific about named third-party tools (ImageTwin, Reviewer Locator, the STM Integrity Hub, similarity checking) and is candid that some AI providers may reuse input data. Yet the retention and training terms of those tools are not disclosed to authors at the same level of detail, and there is no guarantee an author is told that a specific tool flagged their manuscript.
Wiley discloses that its Research Exchange platform uses AI for matching and integrity screening, and its editor guidance helpfully states that a machine-detection flag should not be the primary decision instrument. Its public account of data governance, however, is thin. ACS and IEEE disclose comprehensive pre-screening but leave the AI/ML component of it under-specified. SAGE acknowledges automation and uses image-integrity tools, but offers the least publicly visible detail of the group.
Table 2
Editorial AI tool inventory
| Tool | Publisher | Function | Origin |
|---|---|---|---|
| AIRA | Frontiers | 40+ checks: language, image, plagiarism, ethics, paper mill, COI | In-house |
| Geppetto | Springer Nature | AI-generated fabricated-text detection | In-house |
| SnappShot | Springer Nature | Image manipulation / duplication | In-house |
| Check Integrity | Elsevier | Authorship-change and editorial-COI detection | In-house |
| Research Exchange | Wiley | Plagiarism, paper mill, AI phrasing, identity | In-house / integrated |
| ImageTwin | T&F, Wiley, SAGE | Image duplication detection | Third party |
| iThenticate / Crossref | ACS, Elsevier, T&F | Text similarity / plagiarism | Third party |
| Reviewer Locator | Taylor & Francis | Reviewer matching (WoS data) | Third party (Clarivate) |
| STM Integrity Hub | 35+ publishers | Paper mill, duplication, tortured phrasing | Consortium |
| Paperpal / Papermill Alarm | Frontiers | Paper-mill pre-screening | Third party |
COI = conflict of interest. The inventory lists only tools publishers name publicly; unnamed tools are not represented here.
Table 3
Transparency maturity ranking
| Rank | Publisher | Distinguishing feature |
|---|---|---|
| 1 | Frontiers | Named in-house tool, published checklist, explicit human oversight |
| 2 | Springer Nature | Named tools plus quantitative usage reporting |
| 3 | Elsevier | Clear policy, Check Integrity rollout, responsible-AI principles |
| 4 | Taylor & Francis | Named third-party tools; acknowledges data-reuse risk |
| 5 | Wiley | Platform disclosed, tool map partial |
| 6 | ACS | Narrow scope (similarity checking only) |
| 7 | IEEE | Narrow scope (AI/ML share unspecified) |
| 8 | SAGE | General “we use technology and automation” statements |
Ranking reflects the depth of public disclosure across the four dimensions and is qualitative rather than an audited score.
04Why it matters
Governance, error, and the people at the margins
The stakes are not abstract. Established frameworks already point the way: COPE asks editors and journals to disclose AI use up front; ICMJE recommends that everyone in the editorial process be transparent about which tool was used and for what. Under the GDPR and the EU AI Act, meaningful human oversight is expected to be substantive rather than a rubber stamp, paired with transparent information about the decision and a route to contest it.
The clearest risk is the false positive, and it falls hardest on researchers writing in a second language. A widely cited 2023 study found that AI-text detectors misclassified more than half of TOEFL essays as machine-generated. In 2026, a major machine-learning conference desk-rejected roughly eighteen percent of submissions on the strength of a detector whose flag rate swung between about thirteen and forty-three percent depending on a setting the authors never saw, with no route of appeal. When a superficial feature like an em dash can move a score, article-level transparency and a right to human review stop being a courtesy and become a matter of fairness.
Table 4
Assessment of four possible governance concerns
| Concern | Verdict | Basis |
|---|---|---|
| No meaningful notice at submission | Confirmed | No publisher gives article-level notice of which tools will screen a paper; statements are corporate and general. |
| Manuscript content sent to third parties | Partly confirmed | Third-party tools are used and data-reuse risk is acknowledged, yet some publishers assert “closed” tools; no independent verification exists. |
| Weight of algorithmic output undisclosed | Confirmed | “The human decides” is asserted, but how much a flag raises the odds of desk rejection is not quantified. |
| No AI-specific appeal mechanism | Confirmed | General appeal processes exist, but none is built on a documented right to human review of an AI-driven flag. |
Verdicts are based on publicly available policy statements, not on any independent audit of internal practice.
05A way forward
Four steps toward symmetric transparency
None of this argues against screening. It argues for making the publisher’s side of the process as legible as the author’s is already required to be. A staged path is workable.
- A machine-readable editorial-AI noticeAt submission and in the decision letter, a standard notice listing the tools used, their purpose, and whether each is in-house or third-party, applying ICMJE’s “which tool, which purpose” principle symmetrically to the publisher.
- A data-lifecycle disclosureWhich third parties receive the manuscript, how long data is retained, and whether it is used to train models, closing the reuse risk that at least one publisher already acknowledges.
- A guarantee of human oversight and disclosed weightA binding statement that a flag is never sufficient grounds for rejection on its own, together with published false-positive rates for the tools in use.
- A right of appeal against algorithmic decisionsFor any author affected by an AI-driven flag, a right to human review and an AI-specific route of contest, consistent with GDPR Article 22 principles.
Read together, these steps redefine transparency as a shared, three-way obligation. The current arrangement asks the author for full disclosure while leaving the publisher largely exempt, and that is the imbalance worth correcting.
§Method & limitations
This analysis draws on publishers’ publicly available policy pages, author guides, and press statements as of August 2026, cross-checked against COPE, ICMJE, STM Association materials and independent reporting. The findings reflect what publishers state publicly; this is not an independent audit of their internal practice. Assertions of “human oversight” and “flag, don’t decide” are publisher claims. Usage figures come from annual reports and press releases and are not journal-level. Policies change quickly, so the picture is a snapshot rather than a fixed record.
06Selected sources
- American Chemical Society (2026). Ethical guidelines & AI best practices at ACS Publications.
- Bozkurt, A. (2026). Artificial intelligence in scholarly peer review: ethical considerations, current practices, and future implications. The International Review of Research in Open and Distributed Learning, 27(3), 11–30.
- Committee on Publication Ethics (2025). Emerging AI dilemmas in scholarly publishing (Forum).
- Elsevier (2026). Generative AI policies for journals; Check Integrity expansion.
- Frontiers (2025). About AIRA; broadening AI-driven integrity checks.
- ICMJE (2024). Recommendations: Artificial intelligence.
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779.
- Springer Nature (2026). 2025 Annual Report; AI across publishing workflows; Geppetto and SnappShot.
- STM Association (2025). STM Integrity Hub.
- Taylor & Francis (2025). Misconduct — editorial policies; AI for academic research.
- Wiley (2025). Research Exchange milestone; the role of AI in academic editing.