As agent-type artificial intelligence (AI) rapidly infiltrates university research labs, capable of formulating research hypotheses, writing experimental code, and even designing visuals, establishing refined research ethics standards tailored to specific academic fields has emerged as a top priority in academia.
In the humanities and social sciences, the issue of AI-generated 'fake references' (hallucinations) and the phenomenon of 'ethics washing'—where comprehensive usage disclosures mask ethical concerns—has been highlighted. In STEM and the arts, the key challenge is proving the substantial human contribution and the boundaries of authorship and copyright when AI generates a significant portion of the output.
On November 5, the Education Ministry and the Korea Research Foundation hosted the '2026 Research Ethics Forum' at the Korea Research Foundation's auditorium in Seocho-gu, Seoul, focusing on 'Presenting Research Ethics Guidelines by Academic Discipline in the Age of AI.'
During the forum, experts from the four major academic fields—humanities, social sciences, STEM, and arts—gathered to discuss the limitations of blanket bans on AI usage and declarative guidelines, proposing specific issues and practical solutions that align with each discipline's research methodologies.
“Beware of Overreliance on AI Detection Tools and 'Ethics Washing': Protocols for Originality Checks Needed by Discipline”
In the humanities session, concerns were raised that formal AI usage disclosures could devolve into 'ethics washing,' obstructing researchers' genuine ethical reflection. Jeong Won-seop, a professor at Kyungnam University, pointed out that among 1,068 humanities journals evaluated by the Korea Research Foundation this year, only 65 had adopted AI-related regulations. He emphasized that a simple acknowledgment stating, 'I used generative AI for sentence correction or initial literature searches' cannot absolve researchers of misconduct.
He presented findings from running commercial AI detection programs on submissions to university and graduate thesis competitions, which showed vastly differing detection rates across tools, underscoring that research ethics violations should not be judged solely on imperfect technical detection results.
The issue lies in researchers receiving completed outputs without experiencing the AI's output process, often unaware that they are committing forgery, alteration, or plagiarism. Choi Ye-jeong, a professor at Hoseo University, noted frequent instances of hallucinations where citations appear similar in argument but differ in expression or are entirely incorrect, even when the original researcher's claims are accurate.
As a solution, the need for 'mandatory verification' reflecting the characteristics of each humanities subfield and 'reform of evaluation systems' was proposed. Park Myung-kwan, an emeritus professor at Dongguk University, suggested that literature should verify text through comparison of editions and translations, history should treat AI-suggested sources as 'candidates' for thorough originality checks, and philosophy should institutionalize verification protocols comparing AI-suggested concepts with existing philosophical arguments.
Moreover, Jeong Won-seop, Choi Ye-jeong, and Lee Hyo-jin, a member of the Ministry of Justice's Legal Advisory Committee, agreed that the current 'quantitative research performance evaluation method' based on the number of papers should shift to a qualitative and accountability-focused approach, allowing researchers to invest time in critical verification rather than relying indiscriminately on AI.
“Concerns Over Fake Citations and Confirmation Bias Distorting Policy: Read Original Texts to Break the 'Illusion of Understanding'”
In the social sciences, the potential for AI-generated plausible false prior research and biased data summaries to distort national policy decisions was identified as a core issue. Kim Tae-hwang, a professor at Myongji University, shared simulation results where generative AI was tasked with writing a report based on prior research regarding 'the necessity of effective presidential governance.'
The AI produced a logically coherent report with citations resembling actual academic papers, such as '(Kim Min-seong, 2020),' '(Park Sang-hoon, 2022),' and '(Seo Jin-young, 2024),' but these references were fabricated or cobbled together. Additionally, in a qualitative study involving in-depth interviews with low-income households, the researcher omitted successful adaptation cases to emphasize difficulties, raising concerns about confirmation bias in the rapid summarization and classification processes using AI.
Last year, a new form of research misconduct emerged where AI was used in peer review, with secret instructions embedded in white text within the paper, urging reviewers to write only positive evaluations. Social science experts proposed 'direct verification of primary sources' and 'differential disclosure based on the level of intervention' as solutions.
Jeong Young-jin, a professor at Inha University, cited a study published in the international journal Nature titled 'Illusions of Understanding,' stating that while AI can enhance research productivity, it can also lead researchers to mistakenly believe they understand more than they actually do. He emphasized that simply disclosing hundreds of prompts used in writing a paper is unrealistic, and researchers must read and comprehend the original texts proposed by AI before citing them.
Jeong Jae-yeop, a professor at Namseoul University, also suggested that instead of a vague acknowledgment stating 'AI was used' at the beginning of a paper, the disclosure should be detailed according to the impact on research judgment, such as: ▲simple sentence correction and translation ▲literature search assistance ▲qualitative data coding and classification ▲statistical code writing ▲hypothesis and interpretation suggestions. He also stressed the need to strictly prohibit reviewers from inputting unpublished manuscripts into external AI without permission.
“Even if AI Writes 80% of Fluid Dynamics Code, It Cannot Be an Author: Independent Verification of Physical Laws Required”
In STEM fields, the evolution of AI into multi-agent and swarm agents capable of planning and executing simulations has raised critical discussions around 'authorship' and 'engineering safety and reproducibility verification.' Kwon Jeong-tae, a professor at Hoseo University's College of AI Engineering, warned that while AI is deeply involved in writing computational fluid dynamics (CFD) and finite element analysis (FEA) scripts, as well as detecting anomalies in sensor data, using unverified AI outputs directly in engineering designs could lead to catastrophic safety incidents, such as battery thermal runaway or structural failures, and corporate confidentiality breaches.
Kwon provided specific case studies to illustrate his guidelines. He stated that even if AI writes 80% of the CFD analysis code central to a paper, authorship cannot be granted to AI, and the human researcher who verified the code must take full responsibility.
According to the International Committee of Medical Journal Editors (ICMJE) and the Committee on Publication Ethics (COPE), authors must be able to approve the final manuscript and bear legal responsibility, which AI cannot do.
During the discussion, Jeong Geun-hong, a professor of chemistry at Sogang University, and Kim Jeung-im, a professor of nursing at Soonchunhyang University, outlined specific verification procedures that human authors must establish in engineering research. They emphasized that when AI writes code or interpolates experimental data to generate graphs, human researchers must ensure: ▲independent computational verification based on physical laws such as mass and energy conservation, dimensional consistency, governing equations, boundary conditions, and mesh independence ▲clear differentiation between actual measurements (raw data) and AI-corrected or interpolated values ▲mandatory 'two-person independent review' within the lab to prevent common errors in the same model, and preservation of AI usage logs (model version, random seed, solver environment, etc.).
“Abandon the Illusion of 'AI Co-Creation': Rights Should Be Narrowly Defined by Human Contribution, While Responsibility Should Be Broad”
In the arts and design fields, the need to delineate the boundaries of 'publishing ethics (authorship)' and 'legal rights (copyright)' has come to the forefront due to the normalization of visual image-generating AI. Jeong Jae-hee, a professor at Hongik University's Graduate School of Industrial Arts, conducted a comprehensive survey of ten design journals registered with the Korea Research Foundation (KCI) and found that only four to five, including the Korean Society of Design and the Korean Society of Design Culture, have established separate ethical regulations or announcements regarding generative AI, indicating an urgent need for institutional reform.
In contrast, leading international publishers such as Elsevier and Taylor & Francis strictly prohibit the generation or manipulation of 'primary research result images,' 'clinical and diagnostic images,' and 'graphic abstracts' using AI, requiring that even permitted data visualizations clearly specify the tools, versions, and supervisory scope used in the captions and methods sections to ensure reproducibility.
Moreover, the design field faces challenges such as bias in AI that repeatedly generates stereotypical representations, exclusion of minority groups, unauthorized synthesis of others' likenesses, voices, and styles, and copyright infringement lawsuits related to AI training datasets.
During the discussion, Kim Dong-heon, an adjunct professor at Kookmin University's Graduate School of Design, remarked that in practical accounting, generative AI subscription fees are treated as 'costs' for software like Adobe or CAD, not as 'labor costs' for colleagues. He noted that while the industry treats AI as a tool rather than a collaborator, claiming 'AI co-creation' in papers or works reflects an unconscious defense mechanism to share the responsibility for errors and plagiarism while taking credit for the results.
Kim emphasized that copyright should be granted narrowly based on proven human creative contributions, while research ethics should broadly hold humans accountable for the entire output, including AI-generated portions, reflecting an asymmetrical structure.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.