Academic publishing spent 2023-2025 debating whether AI-writing detectors belonged in the editorial process at all. In 2026, the debate has shifted: several major publishers now run AI-related screening on submissions as a matter of routine workflow, professional bodies have moved from position statements to a joint effort at a shared disclosure standard, and the accuracy problems that made detectors controversial in the first place have not gone away. This guide summarizes what’s actually verifiable about adoption as of mid-2026, distinct from marketing claims made by detection-tool vendors.
What “AI detection adoption” actually means in 2026
“AI detection” gets used loosely to cover at least three distinct practices that publishers are adopting at different rates and for different purposes:
- Text-origin screening — tools that estimate the probability a passage of manuscript text was generated by a large language model (the category most people mean by “AI detector,” and the one with the most documented accuracy problems — see below).
- AI-assisted integrity screening — broader automated checks, some of which use AI/ML techniques themselves, that flag submissions for issues like image manipulation, statistical anomalies, undisclosed authorship changes, or paper-mill signatures. These are not text-origin detectors and are not trying to answer “did AI write this.”
- Disclosure compliance checks — verifying that a manuscript’s methods/acknowledgments section contains a required AI-use disclosure statement, where a journal’s policy requires one.
Conflating these three is a common source of confusion in coverage of this topic. The rest of this guide keeps them separate.
What major publishers have actually deployed
Verifiable, sourced developments as of 2026:
- Elsevier — Check Integrity. Elsevier announced in a March 2026 press release that it had expanded Check Integrity, an automated submission-screening tool, across nearly 2,000 of its journals following a pilot phase. Check Integrity screens for potential breaches of publishing-ethics policy — including unauthorized authorship changes and undisclosed editorial conflicts of interest — and routes flagged submissions to specialist integrity analysts for human review rather than issuing an automatic rejection. This falls into the “AI-assisted integrity screening” category above, not text-origin detection specifically.
- Springer Nature — Geppetto and SnappShot. Springer Nature built an in-house AI tool, internally called Geppetto, that estimates the probability sections of a submitted manuscript contain AI-generated “tortured phrase” or nonsense text, by checking internal consistency section-by-section. Springer Nature has said the tool identified hundreds of fabricated papers shortly after submission since its internal rollout (around 2024). In April 2025, Springer Nature donated Geppetto to the STM Integrity Hub — a shared, cloud-based screening infrastructure run by STM (the International Association of Scientific, Technical and Medical Publishers) — making a version of the tool available to other participating publishers rather than keeping it proprietary. Springer Nature has also deployed a companion tool, SnappShot, aimed at identifying problematic or manipulated images.
- STM Integrity Hub (industry-wide). Rather than every publisher building its own detection stack, a meaningful share of 2026 adoption is happening through this shared initiative: participating publishers submit manuscripts for automated checks (image screening, text-similarity, and increasingly AI-content signals like the donated Geppetto model) through common infrastructure instead of building bespoke tools. This is the more accurate way to describe “industry adoption” than a publisher-by-publisher list — the trend is toward shared, pooled screening infrastructure, not every publisher independently procuring the same commercial detector.
Caveat: figures for exactly how many journals or publishers beyond Elsevier and Springer Nature have deployed comparable screening, and at what specific scale, are not consistently or independently reported as of this writing — treat vendor and trade-press claims about “most major publishers” or specific adoption percentages with caution unless a publisher has made its own announcement, and check current statements directly with a given journal or publisher before relying on a number here.
Policy statements: what publishers and editorial bodies actually require
Separate from detection technology, the policy layer is comparatively well-established and has been stable for roughly three years:
- The ICMJE (International Committee of Medical Journal Editors), the Committee on Publication Ethics (COPE), and the World Association of Medical Editors (WAME) have each held, since 2023, that AI tools cannot be listed as manuscript authors — authorship requires accountability for the work, which a non-human tool cannot bear. COPE’s position statement on this, “Authorship and AI tools,” was published February 2023 and hasn’t substantively changed since; it’s a stable baseline, not a new 2026 development.
- Where publishers do permit generative-AI use in drafting, editing, translation, or literature summarization, the consistent requirement is disclosure: authors must state which tool was used and how, typically in the Methods or Acknowledgments section, and remain fully responsible for the accuracy of anything the tool produced. AI tools generally may not be used to generate or alter research data, images, or figures.
- What is genuinely new for 2025-2026: COPE, together with the World Conferences on Research Integrity Foundation (WCRIF), the International Science Council, STM, and the Global Young Academy, is developing a joint “Global Reporting Standard for AI Disclosure in Research” — informally discussed as the “Vancouver Standard,” tied to the World Conference on Research Integrity held in Vancouver in May 2026. As of mid-2026 this is still in staged public consultation (a second consultation round on what should be disclosed ran through 2026), not a finalized, adopted standard — don’t cite it as already in force.
For a fuller walkthrough of how these obligations apply when you’re the one disclosing AI use in a manuscript, see CASRAI’s AI writing tools hub.
The accuracy and false-positive problem — why adoption is contested
The growth in institutional adoption of integrity screening (Check Integrity, Geppetto, STM Integrity Hub) sits alongside a separate and still largely unresolved problem with the older, narrower category of text-origin AI detectors (tools like Turnitin’s AI writing indicator or GPTZero, which score raw manuscript text for likely AI generation). These tools work by detecting statistical patterns — low “perplexity” and “burstiness” in sentence structure — that correlate with, but do not prove, AI generation. Human-written text that happens to be formulaic (non-native English writers, technical/formulaic disciplines, and heavily edited prose are disproportionately affected) can trigger the same statistical flags. False positives are a documented, recurring issue, not a rare edge case, and no text-origin detector on the market claims forensic-grade certainty.
CASRAI covers this specific problem — why a paper gets flagged, what the evidence actually shows about detector reliability, and what to do if you’re accused of AI use based on a detector score — in a dedicated guide: Why Does My Paper Say “AI Detected”? Understanding False Positives. That guide, not this one, is the right resource if you’re responding to a specific flag. The adoption trend covered here is a separate question from whether any individual detector result should be trusted — a publisher rolling out more integrity screening does not, by itself, mean text-origin AI detection has gotten more accurate; the two categories in this guide’s opening section are worth keeping distinct for exactly this reason.
For a side-by-side look at how individual detection tools compare on the accuracy dimension, see CASRAI’s AI detectors compared comparison page.
What this means if you’re preparing a manuscript in 2026
- Check the specific journal’s current author guidelines before submitting — AI-use disclosure requirements are publisher- and often journal-specific, and policies are still being actively updated as the Vancouver Standard consultation progresses.
- Expect that a growing share of submissions to large publishers pass through some form of automated integrity screening before or alongside peer review — this is now standard workflow at Elsevier and Springer Nature specifically, and increasingly common via shared infrastructure like the STM Integrity Hub, though it is not universal across every journal or publisher.
- Understand that these integrity-screening tools are largely distinct from, and generally more robust than, standalone text-origin AI detectors — a Check Integrity or Geppetto flag is not the same event as a Turnitin AI-indicator flag, and shouldn’t be treated as equivalent evidence.
- If you did use a generative AI tool anywhere in preparing the manuscript, disclose it per the journal’s instructions rather than relying on a detector’s absence of a flag as cover — disclosure obligations under COPE/ICMJE-aligned policies apply regardless of whether any detector would have caught the use.
Frequently asked questions
Do all major academic publishers now screen submissions for AI-generated content?
No. As of 2026, verifiable large-scale deployments are concentrated at a small number of publishers — Elsevier’s Check Integrity and Springer Nature’s Geppetto/SnappShot are the best-documented examples — plus adoption happening through the shared STM Integrity Hub infrastructure. Claims that screening is now universal across major publishers are not independently substantiated; check a specific journal’s current policy rather than assuming.
Is a “Check Integrity” or “Geppetto” flag the same thing as a Turnitin AI-detection flag?
No. Check Integrity is built to catch broader publishing-ethics issues (authorship manipulation, undisclosed conflicts of interest), not specifically to score text for AI generation. Geppetto specifically targets AI-generated “nonsense” or “tortured phrase” text at the manuscript level, using internal consistency checks, and is used internally by Springer Nature (and now, via donation, the STM Integrity Hub) rather than being a general-purpose detector authors can run themselves the way Turnitin’s or GPTZero’s tools are.
Has ICMJE or COPE endorsed a specific AI-detection tool?
No. Neither ICMJE nor COPE endorses or recommends specific commercial AI-detection products. Their guidance is about authorship criteria and disclosure obligations, not about which detection technology publishers should deploy.
Is the “Vancouver Standard” for AI disclosure already in effect?
No, not as of mid-2026. It’s a joint initiative among COPE, WCRIF, the International Science Council, STM, and the Global Young Academy, still in staged public consultation tied to the May 2026 World Conference on Research Integrity in Vancouver. Treat it as a standard in development, not one journals are currently required to follow.
This guide focuses on publisher- and policy-level adoption trends. If you’ve personally been flagged by an AI detector, start with Why Does My Paper Say “AI Detected”? for what to do next. CASRAI discloses that some pages on this site, including the AI writing tools hub and its comparison pages, contain affiliate links to AI-writing and detection tools; that relationship does not change the sourcing standard applied to this guide.
