In 2026, a manuscript over which you have worked for six months will be rejected by an algorithm during the first 30 seconds after submission even before an editor has seen it.
According to an analysis by Wang et al. (2026), published in Learned Publishing (Wiley), 439 highly ranked journals have already implemented formal AI guidelines. In the overwhelming majority of cases, the initial screening is carried out automatically without human involvement. If your AI Probability Score surpasses the publisher’s internal threshold, the article receives a Desk Reject even before the editor opens it.
The biggest problem:
the algorithms cannot differentiate between non-native academic style and machine-generated text. In a follow-up study by Stanford HAI (2026) the average false positive rate for TOEFL essays by Chinese students was 61.3% against 5.1% for texts by American students. Ukrainian, Polish, Romanian and Balkan academics are in the same zone of systemic risk.
This article considers:
exactly how AI detectors work in leading journals, why non-native authors are the hardest affected, and how to prepare a manuscript in such a way that the algorithm allows it to reach the editor. On 29 May, there will be a private webinar where the entire process will be explained in detail.
What changes have been implemented in 2026 regarding publishers’ policies?
The COPE ethic Committee finally established three rules mandatory for all Scopus/WoS journals:
01
AI cannot be an Author. Neither ChatGPT, Claude, nor DeepL carry legal responsibility for fabricated data or copyright infringement and therefore cannot be mentioned in the authorship or acknowledgements as a “co-author”.
02
Transparency is required. If AI was used for translation, proofreading, or stylistic polishing, this must be reported in accordance with the relevant journal’s policy. Concealing a Desk Reject is deemed a procedural breaching.
03
All responsibility lies entirely with the human-authors. Regardless of what the instrument generated, the author always puts the finishing stamp on the product.
How does this compare across different publishers?
Publishers adopted COPE’s principles into their own policies, but with some differences in detail:
Elsevier — permits the use of language support without mandatory disclosure, but stipulates a separate section named “Declaration of Generative AI in Scientific Writing” prior to the list of references if AI was used in the research process.
Springer Nature — provides an exception for AI-assisted copy editing (without disclosure) but insists on documenting the use of LLM in the Methods section.
Wiley — allows AI as a companion in writing, with compulsory disclosure upon submission.
Taylor & Francis — has the most accurate wording: a literal quote is required in the Methods or Acknowledgments.
Minor details in the phrasing make all the difference. Any mistake in the section title or the format of the declaration, and the system interprets this as “concealing the use of AI”.
How does the algorithm “see” you: Perplexity and Burstiness?
In 2026, manuscript submitted to Q1/Q2 journals go through a double filter. This includes a classic similarity index (plagiarism). In addition, there is an AI Probability Score, which is determined by two primary metrics:
- Perplexity. AI is a statistical model that selects the most plausible next word. Therefore, AI-generated texts have low perplexity. A human writer uses unannounced words, complex synonyms and idioms.
- Burstiness. AI writes mechanically writes sentences of similar length and structure, while a living person writes a complex phrase of five lines, followed by a three-word sentence.
If both indicators are low, the algorithm throw up a red flag.
Why does this affects non-native authors?
Academic writing in English is largely based on standard phrases such as “This study aims to…”, “The results indicate that…”, “It should be noted that…”. Whilst such expressions are typical of scientific writing, it also make the text more monotonous and predictable.
For authors B2/C1 level of English, there is an additional risk of using default templates combined with automatic editing through Grammarly, this reduces the variability of the text further. As a result, the model detects a combination of low perplexity and high burstiness, which can be interpreted as a sign of AI-generated text.
A 2023 research by Stanford University, confirmed in a follow-up analysis in 2026, showed that the AUC of detectors drops from 0.89 for texts be native authors to 0.72 for texts by non-native authors.
Essentially, the performance of the algorithms is approximately 20% less effective with texts by Ukrainian authors and tend to make more mistakes that are not in their favour.
Why does this directly apply to Ukrainian scientists?
Here it is worth to be honest. In 2026, Ukrainian authors face three interrelated risks:
01
Linguistic risk
English as a second language and professional proofreading can make the text sound too polished, reduce its natural flow, and may increase the risk of false positives in content detection tools.
02
Reputational risk
In April 2026, Retraction Watch published an analysis by BuyTheBy of over 18 000 advertisements of paper mills in seven countries, including Ukraine. As a consequence, affiliation with Ukraine sometimes increases the risk assessment even before the text is analyzed, creating a systemic bias against honest authors.
03
Procedural risk
Most Ukrainian researchers do not know the exact terms used for AI disclosure by different publishers. An incorrect section title or the absence of a literal quote from the T&F guidelines leads to the Desk Reject for hiding the use of AI, even if the AI was used exclusively for translation.
The combination of these three factors means that the Ukrainian author cannot afford to “wright as usual and see what happens”. The probability of rejection is too high.
What does it mean to prepare a manuscript properly?
Protecting the manuscript from false positives in AI detectors is not about getting around the system. It involves ethical language editing combined with transparent declaration.
What does work?
- Pre-submission AI screening. Passing the manuscript through detectors that the publisher will subsequently use for testing, preceding submission. If the AI Probability Score falls into the danger zone, the text needs to be humanized by enhancing burstiness and adding variation to sentence structure.
- Humanization by native speakers. This is not rewriting but restoring the text’s liveness, while fully preserving the scientific purport and terminology.
- The ready-made declaration tailored to a particular publisher. Accurate formulation, precise placement (Methods/Acknowledgements/separate section), precise name of the tool and purpose of use.
- Reference checking. If an LLM was used for literature search, every DOI should be checked individually. Hallucinations of source are a direct route to retraction.
This is Language Editing in the sense recommended by Elsevier and Springer. These are not grey-area schemes but an officially permitted practice that now demands more expertise than it did a year ago.
Webinar invitation
If you are thinking about submitting to a Q1 or Q2 journal in the following 6 months, you need to understand how the algorithms that make decisions on behalf of the editor work. Otherwise, you might receive a desk rejection for work that does not deserve it.
29 May 2026, 7:00 pm. — closed webinar
We will examine:
- The detailed policies of Elsevier, Springer Nature, Wiley, Taylor & Francis — what exactly to declare and where.
- How automated sorting works and how many seconds your manuscript takes.
- The phenomenon of false positives and why non-native authors suffer first.
- High-profile retractions from 2025–2026 and what we can learn from them.
- A step by step process for ethical manuscript development.
Bonus for all registered participants:
Bonus for all registered participants:
PDF “COPE 2026 manual and templates for statement AI” with prepared phrasing for Elsevier, Springer, Wiley, Taylor & Francis. You can copy this document into your manuscript as early as tomorrow.
