The interaction between AI writing and AI-content detection largely takes the form of a continuing contest between the efficiency and convenience of the former and the regulatory and constraining role of the latter. Image generated by AI

Rapid advances and widespread adoption have made AI an increasingly indispensable tool. New forms of AI-assisted creation can automatically generate text, images, audio, video, and other content, greatly reducing the burden of intellectual labor and enhancing the efficiency of creative work. Yet the growing use of AI writing technologies has also brought a range of problems and challenges, including difficulties in determining originality, potential harm to human cognitive capacities, the spread of superficial and opportunistic academic practices, and heightened risks of plagiarism and AI misuse. These concerns have spurred the rapid development of AI-detection technologies, giving rise to an ongoing contest between AI writing and detection. Since 2023, most Chinese universities have gradually introduced a “dual-check” system for graduation theses, screening simultaneously for text similarity and AI-generated content. This development marks a shift in the contest from technological application to academic governance. While some researchers have focused on improving institutional rules governing AI use, the evolutionary logic of AI technology itself can easily be overlooked. Understanding this trajectory of AI development has therefore become both a serious academic question and a topic of considerable public interest.

Contest between AI writing and AI detection

AI writing models itself on human writing and generally follows familiar patterns of thought and expression. As its capabilities rapidly evolve, the more closely AI-generated text conforms to established writing conventions, the more it tends to resemble high-quality human writing. AI detection, meanwhile, is a specialized technology that addresses the limitations of earlier systems designed solely to identify textual similarity. “De-AIing” takes the opposite approach, reworking textual expression without altering the original meaning, data, viewpoints, or technical terminology. It specifically targets features associated with AI-generated text, often by departing from conventional patterns of expression. The more colloquial, disjointed, or removed from the conventions of polished writing a text becomes, the more readily its AI-detection score can be reduced. AI writing and detection have thus formed a closed technological loop of “generation → confrontation → de-AIing → re-detection.” Detection tools identify features of AI-generated text, users attempt to eliminate these features through rewriting, polishing, and semantic reformulation, and detection technologies evolve in turn to identify new patterns. This cat-and-mouse dynamic drives successive rounds of technological evolution. Most AI-service platforms currently operate on a commercial basis, offering writing and detection services alongside de-AIing and rewriting. As this profit-driven model extends into academic research, its potential negative effects on research and talent development warrant attention.

Current academic discussions of AI writing and detection have focused mainly on specific areas such as technological optimization, the effectiveness of detection tools, and the governance of academic misconduct, with relatively little systematic examination of the deeper tensions, dynamics, and evolutionary logic connecting the two. Rather than simply standing in opposition as “evasion versus detection,” AI writing and detection evolve through continuous interaction. Advances in AI writing spur innovation in detection, while improvements in detection, in turn, drive further advances in AI writing and strategies for evading detection. This process promotes advances in academic-review technologies, but it also gives rise to problems including inaccurate detection, ambiguous evaluation standards, the excessive use of technological means in academic governance, and insufficient theoretical and academic innovation. These issues warrant systematic analysis and closer examination.

At its core, technology can be understood as the processes, methods, or mechanisms through which purposeful human activity is carried out—a means of realizing human desires, needs, purposes, and intentions. Different desires, needs, purposes, and intentions give rise to different forms of technology, which in turn generate contradictions both within and between technological systems. Internal contradictions arise between mutually constraining or opposing components, mechanisms, or tendencies within the same system, such as the seal and pressure-release valve of a pressure cooker, the drive and braking systems of an automobile, or a reservoir’s dam and spillway. External contradictions arise between different technological systems, as seen in missile and missile-defense systems, information-security and information-theft technologies, or aircraft stealth and counter-stealth radar technologies. The distinction between these two forms of contradiction is relative and conditional, and their specific manifestations vary widely. Nevertheless, such contradictions are a direct driving force behind technological evolution and upgrading.

Historically, advances in mechanical, electrical, and information technologies have helped free humanity from arduous physical labor. Today, AI is similarly freeing people from many difficult and complex forms of intellectual labor, while also bringing to the surface a host of tensions involving human desires, purposes, intentions, and values. On the one hand, AI can rapidly perform complex tasks such as drafting texts, conducting literature reviews, developing arguments, and polishing language, substantially lowering the barriers to academic writing and reducing the time it requires. It has thus become an important tool for generating academic content. University students, for example, frequently use AI-assisted tools in academic writing, significantly reducing the time and effort they devote to academic training. This has created an urgent need to reform traditional models of talent cultivation and reshape academic norms. On the other hand, teachers, journal editors, online platforms, academic affairs offices, graduate schools, and other bodies responsible for academic evaluation can no longer rely solely on traditional expert review to reach judgments quickly and accurately, prompting them to turn to powerful AI technologies. This practical need has also contributed to the rapid development of AI detection. The interaction between AI writing and detection therefore largely takes the form of a continuing contest between the efficiency and convenience of the former and the regulatory and constraining role of the latter.

The contest between AI writing and detection reflects the tensions between academic research and academic norms. More specifically, it reveals the growing friction between rapidly evolving technologies for generating academic content and academic governance systems, evaluation standards, and ethical norms that have yet to keep pace. Through continued interaction and escalation, the relationship between the two has passed through three stages: unilateral dominance, two-way confrontation, and normalized coexistence. This evolution has helped foster new paradigms of academic research and driven the intelligent transformation of academic-content review technologies. At the same time, it has created a range of challenges, including unreliable technological detection, distorted evaluation systems, inadequate institutional norms, and weakened academic ethics.

Regulating AI writing and detection

Technological research, development, and application unfold within specific social and historical contexts and inevitably affect people’s thinking, values, interests, and even social order. Historical experience in adopting new technologies and resolving technological contradictions shows that the widespread application of every major technological advance can trigger profound changes in ethics, law, and academic governance, requiring a period of adjustment as emerging technologies challenge established ideas and practices. Building a consensus within academia on responsible AI use is an important starting point, but efforts must also be made to understand and actively guide the healthy development of AI technology, uphold academic standards, preserve the integrity of research, and explore pathways toward greater sci-tech self-reliance and self-strengthening at higher levels. A deeper understanding of AI’s evolutionary logic should inform broad consultation with the scientific, academic, educational, and cultural communities through academic discussions, expert forums, surveys, case studies, and other means, pooling expertise and building consensus. Meanwhile, lessons should be drawn from academic governance practices in China and abroad to establish an integrated system combining technological optimization, institutional improvement, and ethical constraints, with technological breakthroughs as its ultimate objective. Such a system should continue to evolve through academic governance practice, with ongoing revision and improvement aimed at breaking the cycle of confrontation between AI writing and detection.

First, the approach to AI use should shift from “prohibition” to “appropriate and transparent use,” guided by the principle that human judgment remains primary and AI plays a supporting role. Academic evaluation should focus less on whether AI was used and more on how it was used, whether that use was transparent, and whether the resulting work makes a substantive academic contribution. In other words, the focus should be on the appropriateness of AI use rather than on the tool itself. AI-detection results should serve only as preliminary indicators of potential risk, rather than direct evidence of academic misconduct. Statistical features of a text should not be equated with judgments about academic integrity. Instead, evaluation should return to substantive examination of research content, evidence, and logic, with reviewers assessing the validity of research evidence, the logic of the argument, and the authenticity and appropriate use of cited sources. AI assistance with textual expression may be accepted to an appropriate extent, but the thinking and argumentation must remain the author’s own. Particular attention should therefore be given to the research’s problem awareness, academic arguments, research design, use of evidence, and inferential logic.

Second, technical rules governing AI use should be strengthened. A standardized disclosure system should require authors to identify, in the research-methods section, acknowledgments, or a separate declaration, the AI tools and versions used, the specific stages of writing in which AI was involved, and the extent of subsequent human verification and revision. For degree theses, AI interaction and prompting logs should be retained along with successive manuscript versions, making the research and writing process traceable in the event of future disputes or appeals. Rules for identifying and handling academic misconduct should clearly classify AI ghostwriting, deliberate concealment of extensive AI involvement, and the direct use of AI-generated false references or fabricated data as forms of academic misconduct. A clear distinction should also be drawn between “reasonable assistance” and “prohibited ghostwriting,” supported by detailed review procedures, evidentiary standards, and appeal mechanisms to address persistent difficulties in determining improper AI use and resolving disputes over erroneous judgments.

Third, academic ethics education on AI use should be strengthened. Disciplines, departments, and institutions should develop specific ethical requirements for AI use based on their particular circumstances and incorporate them into compulsory research-integrity education for undergraduates, graduate students, early- and mid-career scholars, and researchers more broadly. Case-based education should distinguish among typical examples of “compliant assistance,” “grey areas,” and “academic misconduct,” enabling researchers to understand the relevant boundaries and red lines while discouraging opportunistic attempts to evade detection. Supervisors across institutions should also incorporate AI-use norms into routine research guidance, helping students understand the limits of AI use, reviewing AI-use declarations, examining the research process, and promptly addressing excessive reliance on AI tools or the outsourcing of thinking.

In sum, regulating and guiding the contest between AI writing and detection cannot be accomplished overnight. The task spans numerous fields and stages and requires sustained effort, while the proposals outlined above represent only the tip of the iceberg and may not even address the most pressing issues. Given the continued evolution of AI technology, mechanisms for regular review and revision should be established, allowing the boundaries of AI use, disclosure requirements, and evaluation criteria to be adjusted as needed. The focus of academic governance should shift from detecting textual fraud to ensuring the authenticity and transparency of the research process, thereby maintaining a dynamic balance between harnessing AI to advance scientific research and upholding academic integrity.

Wang Bolu is a professor from the School of Marxism at the University of Science and Technology Beijing.