Talking about the changes that AI has brought to editorial work in the past two years, the first thing Wu Di, editor of Science and Technology Bulletin, thought of was that "the length of papers has become longer and the workload has increased."
Wu Di told reporters from China Youth Daily and China Youth Daily that the previous preliminary review of papers mainly focused on topic selection, format and copy ratio. Nowadays, editors still have to discern whether the language is vague, whether the argument has substantial content, and whether those well-formatted references can find the actual original text. Wu Di has encountered some manuscripts whose language is fluent but the content is empty; the formulaic abstracts are like applying templates; the references appear to be complete, and when searching according to the Digital Object Unique Identifier (DOI), the corresponding document cannot be opened.
Although AI is also helping editors complete repetitive tasks such as language polishing and formatting, Wu Di’s observation has increased the pressure on scientific research integrity review and original material verification. “Relying on duplication detection rate alone is far from enough.”
When AI goes from simple polishing and modification, document retrieval to data analysis, research reasoning and peer review, the academic integrity issues faced by academic publishing are becoming more and more complex: how is a result formed, which links are involved by AI tools, who has verified the facts, and who is responsible for the final published content?
Looking at the "very decent" paper, how many "tricks" are contained in it?
At the 2026 World Young Scientists Summit - Symbiotic Development Forum between Young Scientists and First-Class Journals held in October, Yang Pingping, Asia-Pacific Publishing Director of the British Physical Society Press, showed a set of data: in the past three years, the publisher's submission volume has increased by 55%, but the rate of articles that can be submitted for peer review has dropped from 63.5% to 44.5%, and the number of articles published during the same period was only 28%.
In Yang Pingping’s view, the editorial review process 10 years ago was “not as complicated as it is now,” and many judgments involving academic integrity issues were relatively simple and straightforward; but now, as problems such as paper factories and image manipulation involved in AI are more difficult to identify, the editorial department needs to spend more time determining whether the author’s identity is true, whether the citations are abnormal, and whether the manuscript meets the conditions for submission before sending it for external review. The publishing house has also established a dedicated scientific research integrity team to conduct investigations before and after publication of manuscripts.
Wu Di observed that after the popularization of AI, articles can be mass-produced faster, manuscripts can be washed more naturally, and cross-language rewriting can be more frequent. The routine has changed from "a fake template at a glance" to "it looks like a normal paper." He introduced that common signs include abnormalities in the submission account or email/IP, highly templated abstracts, fake reviewer recommendations, and perfect content but missing original data.
Wu Di said that as it becomes more difficult to identify various academic integrity issues involved in an article, it now relies more on the combination of "system data + manual verification + original materials". For references, the editorial department uses special software to screen, and then double-checks for abnormal author information, incorrect journal names, inaccessible DOIs, and inconsistencies between cited content and the original text. For data and pictures, the original materials, statistical consistency, picture metadata, and whether there are repeated splicing and other issues are checked.
E Haihong, a professor at the School of Computer Science at Beijing University of Posts and Telecommunications, and his team studied the identification of problematic papers from the dimensions of text, experiment, image-text consistency, and citation relationship. The model system established by the team will compare features of scientific images and try to identify images generated by AIGC.
"If you are drawing a flowchart or schematic diagram, AI can be used in this situation. You only need to correctly declare the usage when submitting." E Haihong said that for experimental images from the real world, such as medical images that should be taken from real clinical pathology, if they are generated with one click of AI, then you need to help the editor to identify them.
Gao Yandong, a researcher at the Zhejiang Provincial Institute of Public Policy, said in an interview with a reporter from China Youth Daily and China Youth Daily that the intervention of AI in academic scientific research has not completely changed the types of academic misconduct, but has made the knowledge production process more opaque.
"A piece of text, a picture or even a research conclusion may all be deeply involved by artificial intelligence. The traditional method of identifying problems through text comparison is not enough." Gao Yandong said that academic integrity management needs to shift from simply checking the final results to paying attention to the source of data, the generation process and manual review.
AI rate detection cannot be the only basis for judging academic misconduct
In colleges and universities, as AI generation technology, methods of evading detection, and detection tools are constantly updated, using AI tools to detect the AI rate of student papers has long evolved into a "cat-and-mouse game." In the process of supervising graduate thesis, Li Tong, a professor at Renmin University of China, will also check the AI suspicion rate of the paper in addition to regular plagiarism checks. However, he believes that such tools can only be used as a reference and cannot be directly used as a basis for identification.
Li Tong explained that the detection tool itself is based on a probability model, which is similar to "using magic to defeat magic." He remembers that some students faced a 70% AI suspicion rate and asked: "Which sentence is from AI, and what evidence is there to prove it?" The detection tool can provide a numerical value, but it may not be able to clearly define the source of the specific content.
Wu Di also said that the editorial department currently only uses the AIGC test results as a reference and will not "veto" based on this. When doubts arise, it is more important to go back to the manuscript itself and ask the author to provide original data, pictures, analysis code and AI usage instructions, and then conduct verification based on the reviewer's doubts.
In E Haihong's view, a more detailed distinction should be made between manual writing, which is then edited by AI, and AI generation, which is then manually modified and polished, and full text generation with one click, to help editors review the role that AI plays in the article. She emphasized that technology can undertake detection, matching and clue sorting, but ultimately experts still need to interpret the evidence and make scientific judgments.
Gao Yandong also reminded that it is not appropriate to use AI test results directly as a basis for determining academic misconduct; when sanctions are involved, there should be sufficient evidence and the parties' rights to explanation and defense should be protected.
To judge whether it has crossed the line, you cannot just look at which tool was used or how much text was generated. Gao Yandong believes that more attention should be paid to whether AI has replaced the substantive research work that the author should undertake, and whether it has caused misleading. For example, language polish is usually acceptable, but if the meaning of the argument is changed, the author must re-check; auxiliary analysis of data is also not a violation, provided that the data are authentic, the method is appropriate, and the results can be verified.
These requirements also apply to the editor itself. Yang Pingping introduced that the publishing house is exploring AI-assisted preliminary review in a safe environment. The tool can prompt the match between the manuscript and the journal's requirements and help editors find issues worthy of attention. Some process work can also be automated to improve efficiency, but the tool cannot replace the editorial department in making the final judgment.
AI can be involved in the article, but AI cannot be responsible for the results.
In recent years, the country has successively issued a series of policy documents such as the "Opinions on Promoting the Prosperity and Development of Academic Journals" and the "Opinions of the State Council on In-depth Implementation of the "Artificial Intelligence +" Action", and continued to implement the "China Science and Technology Journal Excellence Action Plan" to coordinate the integration of artificial intelligence into all aspects of scientific research and academic communication. And the "rules" that match it are gradually being established.
On August 19, the selected journals and clusters of the China Science and Technology Journals Excellence Action Plan jointly released the "Adhere to Scientific Research Integrity and Jointly Build First-Class Journals—China Science and Technology Journals Excellence Action Plan to Maintain Academic Integrity Initiative." The third article clearly states: standardize the application of artificial intelligence and clarify the boundaries of use. Strictly standardize the auxiliary application of artificial intelligence technology in scientific research activities and academic publishing, establish and improve the disclosure system for artificial intelligence use, proactively declare the specific uses and degree of intervention of artificial intelligence tools, and ensure clear attribution of academic contributions and traceable responsibilities.
Since this year, many domestic academic journals such as the Journal Center of Xi'an Jiaotong University, Journal of Electronics and Information, Legal Education Research, Journal of Harbin University, and Contemporary Teaching and Research Series have successively released their own AI usage specifications. Reporters from China Youth Daily and China Youth Daily noticed that compared to “one size fits all”, the attitude of the academic community has always been towards “using AI in the sun”.
For example, the Xi'an Jiaotong University Journal Center issued a statement on the standardized use of generative artificial intelligence on August 10, stating that authors can use AI tools in compliance with regulations to carry out auxiliary and transactional work, which is limited to improving research quality and scientific research efficiency; anyone who uses AI tools in the process of academic research, paper writing or revision must fully and truthfully disclose the tool name, version, usage procedures and generated content when submitting. Harbin University issued a statement on the standardized use of generative artificial intelligence in the "Journal of Harbin University" and "Contemporary Teaching and Research Series" on October 15, which clarified that the author must bear full responsibility for the authenticity, originality, and academic standardization of all submissions, and will not accept any generative artificial intelligence tool as the signed author of the paper.
Wu Di also introduced that the "Science and Technology Bulletin" allows reasonable and transparent auxiliary polishing, translation, code debugging, and literature retrieval. It does not allow AI signatures, fabrication of data, pictures, and references, nor does it allow core innovation and full-text writing to be handed over to AI.
“The key is not ‘whether it is used or not’, but ‘whether it is explained truthfully, whether the content is true, and whether the author is responsible.’” Wu Di said that the journal requires authors not to fabricate or conceal the use information of generative AI tools when submitting manuscripts. They should be responsible for the usage and sign an agreement with the author.
E Haihong pointed out at the forum that AI can speed up work such as polishing, retrieval, image editing, and table production, but it cannot replace the author's responsibility for academic opinions. Yang Pingping also said that the British Physical Society Press does not allow AI to be an author, and does not allow reviewers to use AI tools to generate review reports; authors, review experts, and editors should be responsible for the reliability of the paper, the rigor of the review, and the scientific judgment respectively.
Regarding the author's disclosure statement, Gao Yandong proposed that it should explain where the tool is used, what role it plays, and what verification measures the author has taken, instead of just writing a general sentence of "using artificial intelligence."
Gao Yandong believes that the current specific requirements of different disciplines and different journals are still inconsistent, and there is an urgent need to transform the principles into enforceable standards: which uses need to be disclosed, to what extent, and which processes need to keep records. “Disclosure should serve transparency and accountability, not label users.”
In addition, regarding the potential data leakage risks of using AI tools to edit and review unpublished articles, Gao Yandong reminded that these materials may contain subjects’ personal information, medical records, and corporate trade secrets. Even if the name is anonymized, it does not mean that the data cannot identify the individual. The possibility of re-identification after combining the information must also be considered.
Wu Di pointed out that unpublished manuscripts cannot be uploaded to public AI models in their entirety, and information such as unpublished data, core ideas, and the identities of authors and reviewers need to be kept confidential. Editors and reviewers use tools to assist their work, and cannot take unverified AI opinions as the final decision.
Gao Yandong suggested that data should be classified and graded to clarify which materials can be input into the general model, which can only be processed in a controlled environment, and which cannot be uploaded. At the same time, the data retention, training use and cross-border processing rules of service providers should be reviewed. Universities should also provide safe alternative tools and training support so that researchers can use them in compliance with the conditions.
"Artificial intelligence can participate in work, but it cannot be a reason to shirk responsibility." Gao Yandong said, "We can neither regard the use of artificial intelligence as misconduct, nor can we lower the standards of academic responsibility just because it is a new technology."
China Youth Daily·China Youth Daily reporter Wang Jingxuan
