The Growing Challenge of Undisclosed AI in US Scholarly Publishing
Across the United States higher education sector, generative artificial intelligence tools have become deeply embedded in research workflows. Yet policies designed to ensure transparency around their use in scholarly papers are proving largely ineffective at curbing undisclosed assistance. A major 2026 study published in the Proceedings of the National Academy of Sciences reveals that despite widespread adoption of disclosure requirements by academic journals, actual AI-assisted writing has surged without corresponding increases in transparency.
Researchers at Peking University analyzed more than 5,100 journals and 5.2 million papers, finding that approximately 70 percent of journals now maintain some form of AI policy, most commonly mandating disclosure of tool use. However, the presence of these policies shows no measurable impact on the rate of AI adoption. Usage has risen dramatically across disciplines, with particularly sharp increases in the physical sciences and among high open-access outlets.
Landmark Evidence of Policy Shortcomings
The study examined full text from 164,000 scientific publications and focused on the 75,000 papers published since 2023. Only 76 of those explicitly disclosed AI assistance in methods or acknowledgments sections, representing roughly 0.1 percent. This transparency gap persists even as overall suspected AI use climbs. The authors note that disclosure rates have ticked upward slightly over time but remain minimal, suggesting policies have not shifted behavior meaningfully.
US-based publishers and institutions have responded with their own guidelines. Major outlets such as those from Taylor & Francis, Elsevier, and IEEE require authors to disclose generative AI use for language improvement or other assistance while prohibiting AI from being listed as an author. Despite these rules, the data indicate that most researchers are not complying with disclosure expectations.
Surge in AI Adoption Across US Research
Faculty and graduate students in American universities report increasing reliance on tools like ChatGPT and Claude for drafting, editing, and outlining. Surveys from organizations including the College Board and EDUCAUSE show that a majority of faculty have experimented with AI in their own work, while student usage for writing tasks is widespread. Concerns center on critical thinking erosion and originality, with many instructors noting that AI-generated content often evades traditional detection methods.
The phenomenon extends beyond students. Early-career researchers and established faculty alike describe using AI to overcome language barriers, accelerate literature reviews, or polish prose. In fields with high publication pressure, such as biomedical and physical sciences, the incentive to produce more output quickly appears to outweigh disclosure requirements.
University-Level Responses and Limitations
Individual US institutions have issued varying guidance. At places like UC Berkeley School of Law and Harvard, policies emphasize instructor discretion and prohibit undisclosed AI for core writing tasks. Stanford students have publicly discussed opting out of AI entirely in certain programs. Yet enforcement remains difficult, and many faculty report feeling underprepared to guide responsible use.
Broader institutional strategies often focus on academic integrity offices updating misconduct definitions to include undeclared AI assistance. However, detection tools have proven unreliable, leading some departments to shift toward in-class assessments or process-oriented assignments that document human contribution.
Photo by Evgenii Vasilenko on Unsplash
Publisher Policies and the Disclosure Gap
Leading publishers have clarified expectations. Taylor & Francis permits AI for language polishing with disclosure but warns against substituting for core author responsibilities. Similar language appears in guidelines from the Committee on Publication Ethics and individual journals. Still, the PNAS analysis shows these measures have not translated into higher disclosure rates in practice.
High-impact journals with rigorous peer review continue to see suspected AI patterns, including repetitive phrasing and generic structures that detectors flag. The lack of standardized verification methods leaves editors reliant on author self-reporting, which the data suggest is minimal.
Implications for Research Integrity
Undisclosed AI writing raises questions about authorship, accountability, and the authenticity of the scholarly record. When large language models generate substantial portions of text without acknowledgment, readers cannot fully assess the human intellectual contribution. This issue is particularly acute in US research institutions that emphasize originality and rigorous methodology.
Potential downstream effects include propagation of biased or inaccurate content, challenges in reproducing findings, and erosion of public trust in academic output. Early-career scholars worry that undisclosed use could later surface during promotion or funding reviews, creating long-term career risks.
Disparities by Discipline and Demographics
Growth in AI-assisted writing appears strongest in physical sciences and certain open-access venues. Non-native English speakers, including many international researchers at US universities, show elevated rates, reflecting the tools' value for language support. Faculty surveys indicate lower adoption among some underrepresented groups, potentially widening existing inequities in research productivity.
STEM fields report more frequent use for technical sections, while humanities and social sciences see greater application in narrative and argumentative writing. These patterns suggest policies must account for disciplinary differences rather than applying uniform rules.
Challenges in Detection and Enforcement
Current AI detectors suffer from high false-positive rates and can be circumvented. Universities and journals have largely moved away from sole reliance on automated tools. Instead, emphasis is shifting to clear syllabi statements, required process documentation, and honor-code updates that treat undisclosed AI assistance as a form of academic misconduct akin to unauthorized collaboration.
Faculty express near-universal concern that widespread student AI use undermines critical thinking and deep engagement. Many report challenges in distinguishing acceptable assistance from over-reliance, particularly when policies lack specificity.
Emerging Solutions and Best Practices
Experts advocate moving beyond declarative disclosure policies toward verifiable frameworks. Recommendations include requiring authors to retain and share prompts used with AI tools, integrating AI literacy training into graduate programs, and developing discipline-specific guidelines. Some US universities are piloting mandatory modules on responsible AI use for both students and faculty.
Publishers are exploring technical solutions such as watermarking or standardized metadata fields for AI contribution. Collaborative efforts among associations like the Association of American Universities and major publishers aim to harmonize expectations across the sector.
Future Outlook for US Higher Education
As generative AI capabilities advance, the gap between policy intent and actual practice is likely to widen without systemic changes. US research universities face pressure to maintain global leadership in innovation while upholding integrity standards. The 2026 PNAS findings serve as a call to action for reevaluating ethical frameworks and investing in tools and training that support responsible integration rather than prohibition.
Stakeholders across the sector agree that AI is here to stay. The question is whether policies can evolve to promote transparency, accountability, and equitable access while preserving the core values of scholarly inquiry.
