In the AI era, what kind of academic norms do we need?

Editor's Note

From coming up with ideas to conducting experiments and writing papers, AI is deeply intervening in the entire process of scientific research. But behind the efficiency gains, problems such as originality attribution, data "beautification," experimental "absence," information leakage, and paper review are following close on each other's heels, and research integrity and academic norms are facing re-examination.

Narrator: Cao Shuiyan, Associate Researcher at the School of Physics, Nanjing University of Aeronautics and Astronautics

Over the years, I have been engaged in research in micro-nano photonics and related interdisciplinary fields, carrying out theoretical and experimental exploration around two-dimensional materials, micro-nano optical devices, photoelectric detection, and optical information modulation. Since artificial intelligence entered scientific research, my graduate students and I have also been using large models more and more frequently. For consulting literature, sorting out research progress, comparing technical routes, and even discussing whether a new device structure is worth trying, AI can quickly give answers. Sometimes I also ask AI directly: "How else can this direction be innovated?" A few seconds later, it may list five or six "research ideas": switch to another two-dimensional material, introduce a new degree of freedom for light field modulation, combine polarization with photoelectric detection, or further extend to functions such as optical logic and information encoding. Looking at the text alone, some of these ideas are even quite inspiring.

Thus, a problem that rarely arose before has emerged: if research concepts, innovative ideas, or even topic frameworks initially come from AI, when they ultimately form research outputs such as papers and patents, how should originality attribution be determined, and how should the boundary of contributions between researchers and AI be defined?

In my view, at least judging from my own experience in micro-nano photonics research, judging the originality attribution of the final output merely by who first proposed a research idea is to understand scientific innovation too simplistically. In our micro-nano photonics research, we often start from an initial device concept. For example, introducing micro-nano structures into two-dimensional material optoelectronic devices, changing the local light field through structural design, and then using the photoelectric response of two-dimensional materials to achieve new detection or modulation functions.

Similar ideas can now be generated by AI a dozen at a time. But there is still a long distance between an AI-generated idea that "looks novel" and truly completing an original piece of research. Some research ideas seem inspiring at first glance, but after actually carrying out theoretical analysis and numerical simulation, one finds they are not feasible: some ignore necessary boundary conditions; some use material parameters that are divorced from reality; some have a certain effect in theory, but the signal magnitude is too small for experiments to measure; and some so-called "innovations," only after further literature searches, turn out to have similar recent work, so their originality itself needs to be re-examined.

I increasingly feel that what AI can generate are research ideas, not scientific innovations that already hold. Whether a research idea truly has innovative value still requires researchers, in light of existing research, to continuously test its originality and scientific feasibility through literature searches, theoretical analysis, numerical simulation, and even experimental verification. Taking micro-nano photonics research as an example, a device may show very ideal spectra, modulation depth, or polarization response in theoretical calculations, but after it is actually fabricated, factors such as material thickness, processing errors, and electrode contact may all cause the experimental results to differ greatly from the initial conception, or even directly negate the initial concept.

And in this process, new scientific questions may also emerge. AI may have participated in generating the initial concept, but when judging originality attribution, one should not only look at "who first proposed this idea," but should pay more attention to who made the key scientific judgments during the formation of the output and who completed the substantive creative work. This is perhaps the key to clarifying the boundary of contributions between researchers and AI.

This judgment also extends to authorship of outputs and intellectual property. Even if AI participated in generating the initial research approach, one cannot simply conclude on that basis that AI should become an author of a paper or an inventor of a patent. The scientific understanding or technical solution ultimately formed crystallizes the results of researchers' sustained judgment, verification, and creation. More importantly, research outputs mean not only "contribution" but also "responsibility." After a paper is published, if there are errors in theoretical derivation or the experimental results cannot be reproduced, it is ultimately the researchers who need to explain; when facing questions from peers, they also cannot evade responsibility by saying "this research idea was generated by AI." The ones who ultimately stand behind the output and are responsible for it are still the researchers.

In the future, a scientific concept may very likely arise from repeated human-computer interaction, and "who the initial idea came from" may become increasingly difficult to trace, but the source of an idea cannot simply be equated with the originality attribution of the output.

Narrator: Liu Xiaoming, Associate Professor at the School of Electronic and Information Engineering, Xi'an Jiaotong University

My research direction is spatiotemporal anomaly detection and AI-generated content (AIGC) detection. The research work can be divided into two parts: one is training models to identify anomalies from data, and the other is determining whether text, images, or videos are generated by AI. The core premise shared by both research directions is data authenticity. In recent years, while serving as a reviewer for academic journals, I have frequently received submissions in which the methods, figures, and metrics show no obvious problems, but the authenticity of the data is questionable.

The first thing that aroused my vigilance was a submission in the field of spatiotemporal anomaly detection: the authors used traffic flow data from a certain city's road network for their research, and the smoothness of the resulting curves was close to that of textbook examples. But everyone in the field knows that real traffic flow sequences inevitably fluctuate, and various factors such as sensor errors, clock deviations, traffic control, and weather changes will all alter road network characteristics. I asked the authors to provide the original sampling frequency and the method for handling missing values, but they did not respond directly.

Traditional data misconduct mainly involved tampering with real data, which often leaves traces; the more common situation today is real data being "reconstructed" by AI—after denoising, imputing missing values, and smoothing jumps, individual data points appear reasonable, yet the overall distribution can no longer reflect the true situation.

If the first category is "makeup" at the data level, the second category appears at the experimental design level and is even more widespread: in some submissions, the F1 score of anomaly detection models (a performance evaluation metric ranging from 0 to 1, with values closer to 1 indicating a better model) is as high as 0.99, comprehensively surpassing existing methods on multiple public benchmark datasets. When reviewing, I focus on three points: how the training and test sets are split, whether results from multiple random runs are reported, and whether ablation experiments were completed. Some submissions cannot withstand scrutiny. Such problems do not necessarily stem from subjective malice; they are often due to researchers' insufficient understanding of research norms, but AI tools have objectively lowered the threshold for "unconscious violations" and made it harder for reviewers to distinguish negligence from deliberate fabrication.

The first two categories still fall under methodological misconduct, while the third category directly undermines the legitimacy of research and is also the issue that worries me most. Training AIGC detection models requires a large amount of genuine human-created content as negative samples. I once reviewed a submission in which the authors claimed to have collected content from real users on a certain platform, but the text style, punctuation habits, and distribution of typos were highly consistent, matching the typical characteristics of AI batch generation. After it was returned for revision, the authors admitted that, to expand the dataset, they had used a large model to synthesize some samples and had not disclosed this in the article. The absurd part is that the goal of the study was to identify AI-generated content, yet the training data came from AI; no matter how high the metrics are, it is merely self-circular validation.

My doctoral supervisor, Academician Guan Xiaohong, often says that research should come from the physical world and return to the physical world, and that the usability of a model depends on its operational stability under real working conditions. In power systems, an anomaly that has been smoothed away may correspond to an undetected safety accident. Professor Shen Chao, who has worked with me for many years, also proposed that the safety boundary of AI is delineated by labeling failure scenarios, and that a paper willing to clearly state the limitations of its method's applicability is far more valuable than five articles that report only positive results.

But today, AI has reduced the cost of "beautifying" data to almost zero—a single prompt can complete data organization, result reconstruction, and the generation of supporting explanations, and even the "discussion of limitations" can be ghostwritten by AI. The risk of academic misconduct is thus multiplied: a casual data optimization is enough to invalidate the conclusions of an entire paper; if such models flow into engineering sites, the losses caused by false positives and missed detections will ultimately be borne by frontline operations and maintenance personnel and even the general public.

Narrator: Zhou Mengran, Associate Researcher in the Department of Mechanical Engineering at Tsinghua University

Once, a doctoral student in my group used existing experimental data and literature data to build a model for predicting the mechanical and corrosion resistance properties of a magnesium-lithium alloy after friction stir processing. Based on modeling and prediction of parameters such as tool rotation speed and processing speed and the final material properties, the model quickly produced a relatively clear "optimal process window": one parameter combination scored highest, had the best predicted strength, and also had the lowest material corrosion rate. During a project discussion, the doctoral student excitedly suggested to me: "Professor, since the prediction results are already so clear and definite, we only need to do two or three groups of 'optimal parameters' for validation afterward, and the remaining experiments can be skipped. This saves materials and machine time and also allows the paper to move forward faster."

My work involves using friction stir processing to conduct research on welding and additive manufacturing of light alloy materials such as aluminum and magnesium. The physical process looks very simple: a stirring tool head rotates, inserts into the material, and advances along the weld seam, ultimately enabling the material to be joined under "non-melting" conditions, but in fact the internal situation of the material during this process is far more complex than imagined—temperature changes, strain gradients, strain rates, material flow, dynamic recrystallization, texture evolution, and interface states are coupled with one another, and any local change may affect the properties of the final material obtained.

For this reason, I explained to the whole group: I do not reject AI. On the contrary, I believe it will become an indispensable tool for future scientific research. It can help us narrow the experimental scope, optimize plans, propose hypotheses, and even tell us which places are most worth prioritizing for validation. But at present, AI "based on large language models" can only give "what is most likely to happen based on theory and others' results," whereas experiments need to answer "what actually happened and how it happened." If scientific research only validates the answers the model is most confident about, it may appear that efficiency has improved, but in reality it also deletes the places most likely to produce new discoveries.

We both used the model's inferences to achieve a significant improvement in the comprehensive performance of the magnesium-lithium alloy within the "optimal process window" given by the model, and insisted on producing samples with extreme parameters that the model considered to have "low potential value" and that no one had ever tried.

Facts soon taught us a lesson. For one set of parameters that AI rated highly, the material surface looked intact, but cross-sectional observation revealed that local material had occasional flow instability defects; on the contrary, for another set of extreme parameters outside the model's prediction range, the internal microstructure of the material instead showed unprecedented uniformity and density, and the mechanical and corrosion behavior were also far superior to the model's "prediction." We compared all publicly reported literature at the time and found that we had "unintentionally" prepared an ultralight magnesium alloy with the best "specific strength-corrosion resistance" performance.

Faced with this breakthrough phenomenon, we carried out microstructural observation, mechanical property testing, and material flow analysis in quick succession, tracing the cause step by step, and ultimately found that what determines material performance is not merely the "level of heat input and speed of material flow" that are simplified in the model, but rather a completely new mechanism of competitive "deformation—precipitation—diffusion" among multiple components in the material under complex "thermal—mechanical" coupling.

More importantly, that experimental point that was originally meant to be omitted instead became the most critical research direction and clue in my subsequent research. Over these years, my team and I have continued to study this new mechanism behind the anomalous result, readjusted our processing methods and research approach, and gained a deeper understanding of the microstructure evolution and material flow mechanisms in friction stir processing. The related research has thus made progress beyond expectations and has further expanded to the additive manufacturing of high-performance biomedical magnesium alloys and ultra-large magnesium alloy structures.

Looking back now, I often feel that this was in fact a process of "truth defeating laziness."

The so-called "laziness" does not mean that researchers cannot use tools to improve efficiency. Scientific research certainly needs to improve efficiency, and AI can absolutely help us reduce a large amount of repetitive labor. What truly deserves vigilance is quietly turning "improving efficiency" into "omitting empirical verification," and then packaging model predictions as facts that have already occurred.

AI prediction can reduce blind experiments, but it cannot replace necessary experimental verification. AI is good at finding patterns from existing knowledge, so it is more likely to tell us "what usually happens." But scientific progress often comes precisely from one question: "Why is this time different?"

Without real experiments, anomalies may be filtered out as noise, and the unknown may be "predicted away" in advance by existing patterns. In the end, the models become more and more beautiful, the curves in papers become smoother and smoother, yet human scientific research is locked into an internal cycle within existing knowledge.

Risk of scientific research information leakage must be guarded against

Narrator: Deng Shenzhen, Associate Professor, College of Chemistry and Chemical Engineering, China University of Petroleum (East China)

In 2018 I entered Nankai University to pursue a doctorate, with a research direction in aqueous zinc-ion batteries. At that time, AI had not yet entered daily life, and the rhythm of research life was simple: dormitory, cafeteria, laboratory—three points and one line. During the day I synthesized materials and measured data; at night I read literature, organized results, and wrote papers. Literature searches required flipping through items one by one, and the reference list of an important paper also had to be checked entry by entry, following the citation chain downward. Writing English papers was even more time-consuming, requiring extensive reading of top international journal literature, noting good sentence patterns, and building a corpus. When it was truly time to write, it was still not easy; revising a paragraph a dozen times was common, and it was normal for a paper to take five or six months from first draft to submission. At that time I felt that research was supposed to be like this—slow and clumsy, but every step solid. All data circulated within the research group.

In 2022, I went to Hong Kong for postdoctoral research. That year ChatGPT became popular, and people in the group began using it to polish English and search literature. At first I did not think much of it, until one group meeting when a junior fellow student said he had "fed" raw data to AI for analysis, and AI not only organized the patterns in the data but also generated a discussion framework. I asked him what he had uploaded, and he said it was all the test results from the past six months, not yet organized or analyzed, and certainly not published. A thought suddenly came to my mind: after these data are uploaded, do they still belong to us? I could not help asking him: "Aren't you afraid that after these data are uploaded, the platform will share them with others?" He was silent for a long time and did not answer.

After that day, I specifically looked at the service agreements of several mainstream AI platforms and found that almost all contained one sentence: content uploaded by users may be used to optimize services. No one could say clearly what "optimize" specifically meant. During that period I kept thinking: if my junior fellow student uploaded the data and these data were used to train models, then they would no longer belong only to us. If a peer happened to ask AI related questions at some point, would AI "share" these data?

In 2024 I returned to the research group where I had studied for my bachelor's and master's degrees and began supervising graduate students. Unexpectedly, the students in the group had done the same thing. One student organized nearly a year of experimental data into a document and was preparing to upload it to an AI platform to "help analyze." I stopped him: "If these data are leaked by the platform to others, what will you do if the research results are replicated by others or published by them first?" He was stunned and said something that left a deep impression on me: "I just wanted to get results quickly."

At that moment I realized that this was not one person's problem, but a common mentality among a generation of researchers—using the fastest way to process data, yet never thinking about what would happen after the data leave the laboratory. From then on, I set a rule in the research group: it is strictly forbidden to upload unpublished experimental data, process parameters, or research ideas to public AI platforms. AI can be used for work that does not involve core data, such as searching literature and formatting references, but anything involving unpublished data and experimental plans is uniformly prohibited.

Several later leak incidents proved that this worry was not excessive. In July 2025, the Ministry of State Security disclosed that a certain researcher had uploaded core data to an AI application, causing the leakage of classified information. In September of the same year, the ChatGPT Deep Research function was exposed to have a zero-click vulnerability: an attacker needed only one malicious email to steal sensitive data without the user's knowledge. Looking at this news, I thought of that student. If I had not stopped him at the time, where would those data be now?

Paper review needs new standards

Narrator: Cai Xiaolin, Associate Professor, College of Resources and Environment, University of Chinese Academy of Sciences

In the course of long-term participation in peer review in the environmental field, we have increasingly clearly felt that generative artificial intelligence is entering every link of scientific research. From literature review and language editing to code writing, data analysis, and model prediction, AI has improved the efficiency of information processing and also expanded the ways researchers pose questions and verify ideas. But for academic journals, the challenges brought by this change are not limited to whether a paper looks as if it was written by AI. More importantly, as AI gradually participates in knowledge production, can the original disclosure methods, review standards, and boundaries of responsibility continue to support academic trust? The focus of discussion should shift from whether AI is used to how, after AI intervention, the credibility of research is proven and responsibility is traced.

From review practice in the environmental field, a paper with complete structure, fluent expression, and full figures and tables can be produced faster, but the completeness of the text is increasingly less able to represent the solidity of the research itself. Environmental research usually involves sampling, experimental analysis, instrumental characterization, statistical modeling, and mechanistic explanation, and conclusions are also constrained by factors such as regional characteristics, temporal scale, and experimental conditions. AI can rapidly connect concepts and supplement arguments, and may also use seemingly reasonable expressions to conceal unclear data sources, insufficient description of conditions, or broken chains of evidence. Editors and reviewers therefore need to ask: Where do the data come from? Can the methods be reproduced? Do the inferences go beyond the boundaries of the evidence?

The difficulty of review stems to a large extent from the fact that AI amplifies the information asymmetry inherent in peer review. Reviewers can see the text and materials submitted by authors, but may not be able to reconstruct the entire process by which data were processed and conclusions formed. Judging AI traces solely by wording may wrongly harm normal writing and also cannot identify manuscripts that have been manually polished but whose research process is unreliable. Review standards should return to scientific evidence itself and use a more transparent research process to reduce the difficulty of judgment.

The difficulty of definition requires journals to update their AI disclosure systems. The current common binary declaration of "used" or "not used" in journals is too general; the degree of disclosure should match the impact of AI on research judgments. If AI is used only for routine assistance such as language polishing and formatting adjustments, a brief explanation may suffice; if it participates in data cleaning, parameter selection, model construction, figure generation, or literature synthesis and affects interpretation of results, the specific uses and manual verification methods should be explained; if it further participates in experimental design, simulation or synthetic data generation, key inferences, and the formation of core conclusions, more sufficient input information, verification processes, and reproducible materials should be provided. AI helping researchers propose hypotheses should not be uniformly regarded as crossing a boundary. What truly warrants vigilance is concealing substantive use, replacing real evidence with generated content, or directly writing unverified output as scientific conclusions.

For research involving AI, the focus should be on checking whether the data are real and traceable, whether model training and validation are clear, whether key processing steps can be reproduced, whether uncertainty is truthfully presented, and whether conclusions are supported by independent data or experiments. In environmental research, attention should also be paid to whether model results correspond to real environmental conditions and whether patterns obtained from local data have been over-extrapolated. The deeper AI's impact on the research process, the greater the author's responsibility to provide sufficient transparency and verifiable materials.

Peer review itself also needs new norms. AI can help reviewers organize questions and check expression, but it cannot replace their independent judgment on innovation, methodological reasonableness, and strength of conclusions. Manuscripts under review usually contain unpublished data, ideas, and collaboration information; without permission, they should not be uploaded to open tools lacking confidentiality safeguards. Even if a journal permits the use of AI assistance under secure conditions, reviewers should still be responsible for their final opinions, and editors should also prevent template-like reviews that are fluent in language but empty in content from affecting decisions.

(Project team: Our reporters Cui Xingyi and Chen Haibo, and our correspondent Wang Yuhan)

1; Live preview | Paper mill prevention and AI academic ethics guidelines

2; New model ignites AI circles: it does not generate text but makes rapid decisions specifically for machines

8; Nobel laureate Leavitt: AI will be applied to every link of drug R&D

2; Paper topic inspiration: "Sustainable Food" research direction | MDPI Sustainability

3; FDE From lesson preparation to grading, AI is reshaping the classroom, but there is one thing it cannot do

4; FDE Under the technology carnival, who is bearing the cost of AI education?

5; Paper topic inspiration: "Sustainable Urban and Rural Development" research direction | MDPI Sustainability

6; Paper topic inspiration: Cerebrovascular diseases (1) | MDPI Brain Sciences

7; New framework combining AI and multi-source data provides a new solution for agricultural remote sensing mapping

8; Paper topic inspiration: "Sustainable Development: Agricultural and Forestry Waste Treatment and Reuse" research direction | MDPI Sustainability

1; Western cryosphere accelerating "degradation": glacier melting sounds the alarm for high-mountain safety

2; The 6th National Agricultural Science and Technology Development Strategy Think Tank Alliance Conference held in Lanzhou

3; The 2026 Nobel Prizes are about to be announced, and authoritative experts will provide you with a comprehensive interpretation!

8; He became the first person in the national medical field to earn a doctorate without a paper

Famous doctoral dissertations 163: Nobel laureate Peter Doherty's doctoral dissertation, University of Edinburgh, 1970

My complex about famous universities and my experience at a famous university

More>>