09/07/2026

As AI Becomes Embedded in Research, Can Transparency and Academic Integrity Keep Pace?

by Zahra Niazi

Image credit: Author using ChatGPT

In late 2022, when OpenAI made its generative artificial intelligence (GenAI) model, ChatGPT, publicly available, the response of editors and instructors across research and academia was one of cautious optimism. Although they recognised the potential of AI in improving research efficiency, a major concern was how to ensure transparency about the extent and nature of AI assistance in research and whether the researcher using such assistance had complied with the principles of responsible research conduct. For some, problematic AI use was limited to relying on AI-generated reference material without verification. For others, it was broader, including practices such as overreliance on these tools for generating ideas without questioning them, which they argued risked making the knowledge base not only more homogeneous, but also more biased.

AI Use, Disclosure, and Detection

Four years on, AI is increasingly being embedded into research processes. According to Wiley’s ExplanAItions 2025 study, global AI adoption among researchers increased from 57 per cent in 2024 to 84 per cent in 2025 (Wiley, 2025). Yet, questions of transparency and academic integrity remain. A study published in 2026, examining 5,114 journals and over 5.2 million papers, found that 70 per cent of journals had adopted AI policies, particularly disclosure requirements (He & Bu, 2026). However, when examining a subset of 75,000 papers, authors found that only 0.1 per cent had disclosed AI use.

In addition, while detection tools for determining the percentage of AI-generated text have proliferated, they cannot reliably assess the extent of AI use, let alone identify the manner in which it was used. A recent study found that Turnitin, one of the most widely used detection tools worldwide, accurately identified fully human-drafted content, and its AI-detection score increased as the actual proportion of AI-generated text rose (Atamhenwan, 2026). Still, there were instances where it overestimated AI use at lower levels, while its accuracy declined when AI-generated text was humanised.

Then comes human judgement and the role of training in better equipping instructors, editors, and reviewers to make informed judgements about content. However, human judgement, too, has its inherent limitations; one study found that although feedback-based training improved participants’ accuracy in distinguishing AI-generated text from human-written text by about 17.5 per cent compared to the no-feedback group, the feedback group’s overall accuracy was still 65.1 per cent (Milička et al., 2025).

Combining Detection, Disclosure and Human Judgement

In real-world settings, solutions are rarely perfect, and the key is to adopt the best of the available options. Informal exchanges with editors, instructors, and reviewers have suggested that such an approach may involve using all three methods in conjunction. Detection tools can serve as an initial screening tool, although they cannot provide conclusive evidence of misconduct; human judgement can provide context; while disclosure requirements can act as an unspoken moral check, even if actual disclosures remain limited. They can also allow individuals who may have used AI for minor assistance, such as improving prose, to clarify their position before their work is processed through detection software; however, mandatory disclosure should be required only for substantial AI assistance, defined by an organisation or an instructor, while other forms of assistance may be disclosed voluntarily.

From Restriction to Responsible AI Use

Above all, accompanying these measures should be policies that promote legitimate AI use in research. Outright restrictions or policy ambiguities not only deprive organisations of the benefits of these technologies but can also become drivers of problematic AI use. In my professional capacity, for instance, I recently delivered a lecture on the use of technology in research to graduate-level students working as interns at the organisation. During the input session, several stated that research demands had grown with the proliferation of AI tools. However, AI-use policies at their institutions were either overly restrictive or ambiguous, leaving them uncertain about which outputs could at least be drafted with AI assistance if the ideas remained their own. Some stated that, unable to cope with these growing pressures, they resorted to over-relying on AI tools and subsequently concealing their use. A further reinforcing factor, according to them, was the tendency of instructors to view all polished work, including that produced through sheer hard work, with suspicion, which discouraged genuine independent effort.

What constitutes those policies will, however, differ from one organisation, department, and even one course or assignment to another. At the university level, for instance, students are still in the learning stage, so greater caution may need to be exercised when permitting AI use, but this should not lead to outright prohibition. For example, if the objective of a research assignment is to assess students’ ability to perform and interpret statistical analysis using software, and the assignment requires the mandatory submission of datasets and statistical outputs, followed by presentations, AI tools may be allowed to assist with report drafting. In the coming times, as the capabilities of AI tools continue to advance, it is also plausible that providers could begin placing greater limits on their free plans, and, to ensure equitable access for all, institutions may need to provide subscription-based accounts.

However, a difference in critical thinking ability is already evident between professionals who received their education before the advent of AI and students today, or even younger professionals, who will become the problem solvers of tomorrow. To preserve critical thinking ability, traditional assignment formats may no longer work and may need to be redesigned, for example, by incorporating simulation- or scenario-based tasks. However, this approach must be balanced with allowing legitimate AI use where deemed appropriate to avoid placing undue pressure on students, which can itself contribute to cognitive overload.

Likewise, think tanks are required to produce timely analyses on a wide range of issues. To enable more substantive intellectual work, AI-assisted drafting may be permitted for certain types of outputs, such as the compilation of reports after core analysis has been completed and defended, routine briefs, critical summaries, or even a select number of articles under tighter deadlines, provided that the final output reflects originality and depth of thinking, sound analysis, and factual accuracy.

In addition, researchers should be acquainted with the most efficient AI tools for purposes such as summarising and extracting data, compiling references, fact-checking, or identifying research gaps, among other things. At the same time, they must understand their limitations, the importance of human verification, and issues related to uploading sensitive or unpublished information on AI platforms.

The key is to combine human intellectual agency with the capabilities of AI tools to ensure that the knowledge base in the AI era becomes more credible, diverse, intellectually rich, and innovative than it was before the advent of AI.

Keywords: Transparency, Academic Integrity, OpenAI, Research Integrity

References

Atamhenwan, L. E. (2026). How are combinations of human-written words and LLM-generated words by ChatGPT, Copilot, Gemini and Grammarly detected by Turnitin? Education and Information Technologies. https://doi.org/10.1007/s10639-026-14049-2

He, Y., & Bu, Y. (2026). Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing. Proceedings of the National Academy of Sciences, 123(9), Article e2526734123. https://doi.org/10.1073/pnas.2526734123

Milička, J., Marklová, A., Drobil, O., & Pospíšilová, E. (2025). Learning to detect AI texts and learning the limits. PLOS ONE, 20(10), Article e0333007. https://doi.org/10.1371/journal.pone.0333007

Wiley. (2025). ExplanAItions 2025: Key insights. https://www.wiley.com/content/dam/wiley-com/en/pdfs/explanaitions/explanaitions-mini-report-2025-final-2-1.pdf

Zahra Niazi is a Research Associate at the Centre for Aerospace & Security Studies (CASS), Islamabad, Pakistan.