AI could open academia to researchers long constrained by language, background and limited access to elite networks. It could also flood journals with superficially flawless yet increasingly similar manuscripts and push an overloaded peer-review system beyond capacity. At the 2026 Responsible Research Summit, Professor Anna Morgan-Thomas and Professor Dorota Dobija debated what authors, reviewers, editors and universities must now be accountable for.
This article is an edited account of presentations by Professor Anna Morgan-Thomas and Professor Dorota Dobija and the discussion moderated by Professor Desislava Dikova at the 2026 Responsible Research Summit. The contributions have been shortened and edited for clarity while preserving their meaning and the distinct positions of the speakers.
When a tool begins to shape the claim
Who is responsible for a scholarly claim developed with the help of artificial intelligence? What uses of AI are acceptable, how should they be disclosed, and can reviewers responsibly use the same tools? Professor Desislava Dikova of Kozminski University and WU Vienna opened the panel with questions that reach far beyond the working practices of individual researchers. They concern the rules governing the entire system through which knowledge is produced, assessed and legitimised.
As successive stages of research become automated, more than the pace of publishing is changing. Originality, diversity of perspective, the confidentiality of research materials, the quality of peer review and the ability to identify who is genuinely accountable for a conclusion are all at stake.
Scholarly publishing was never a level playing field
Professor Anna Morgan-Thomas, Professor of Digital Management and Innovation at the University of Glasgow’s Adam Smith Business School, described herself as a “romantic optimist” about AI. Yet she began not with a catalogue of benefits but by challenging the idea of a pre-generative-AI golden age. Today’s concerns are legitimate, she argued, but they should not obscure the flaws that were already embedded in academic publishing.
Success in leading journals has long depended on more than the quality of the research. Access to influential academic networks, suitable mentors and co-authors, familiarity with editors, language and research support, and protection from excessive teaching and administrative work all matter. An analysis published in Research Policy, cited during the presentation, found that 51.2% of papers in 18 top-rated management journals involved at least one author from the world’s top 100 business schools.
AI could reinforce that concentration, but it could also expose its mechanisms and lower some barriers to entry. For scholars writing in English as a second or third language, tools that support editing need not be a shortcut. They can level the field in a system that often rewards not only good science but also fluency in a recognisable academic style.
We need to hear good science, not just well-presented science, Professor Morgan-Thomas argued.
Accountability for claims, not tools
This diagnosis led to her central proposal: governance should focus less on whether a researcher used a particular tool and more on whether the researcher can defend the claims made in the publication. A ban on AI use would be difficult to enforce, while disclosure could become a merely performative exercise if it were not accompanied by substantive accountability.
In her article “Beyond AI disclosure: Claim accountability and responsible research in scholarly publishing”, Professor Morgan-Thomas argues that authors should remain accountable for their ability to explain, justify, reconstruct and revise every claim shaped with AI. During the panel, she identified two thresholds for responsible use.
The first is confidentiality. Unpublished manuscripts, participant data and protected materials should not be entered into open AI systems. The second is judgement: the researcher must understand the reasoning presented and retain the capacity to challenge it. If an author cannot explain where a conclusion came from, examine its foundations or revise it in response to scrutiny, they should not put their name to it.
We need to be able to defend what is being produced, she concluded
Twenty manuscripts before the first coffee
Professor Dorota Dobija, Head of the Department of Accounting and Corporate Governance at Kozminski University, approached the same transformation from the perspective of an editor and reviewer. She offered a simple scenario: an editor opens their inbox in the morning and finds twenty new submissions. Every manuscript is fluent, well structured and supported by a plausible rationale. Yet the editor cannot see the extent to which AI shaped the research question, literature review, method, data analysis or conclusions.
The problem is not simply one of detecting fraud. Editors must assess the outcome of a process that is increasingly invisible, while existing publishing systems were designed primarily to identify textual similarity. They cannot reliably reconstruct how an argument was developed or determine where assistance ends and the substitution of scholarly judgement begins.
Scale makes the problem more acute. Twenty manuscripts may require forty reviewers to agree to assess them. Professor Dobija recalled occasions on which she had to send twelve invitations to secure two reviewers for a single paper. If AI allows the volume of polished submissions to grow faster than the scholarly community’s capacity for careful reading, peer review could become the critical bottleneck in the system.
Better language, more similar papers
Research on biomedical publishing already indicates a sharp change in the language of abstracts following the spread of generative AI. The authors of a study of more than 15 million abstracts estimated that at least 13.5% of the abstracts published in 2024 may have been processed with large language models, with substantially higher rates in some subcorpora. This is a statistical inference based on shifts in vocabulary, not proof that the identified texts were written entirely by AI. It nevertheless indicates the scale of the transformation.
For Professor Dobija, the greatest risk is not simply that more papers will be produced. They may also become increasingly homogeneous: similarly structured, dependent on the same readily accessible sources and drawn towards safe, predictable conclusions. Models learn from the existing literature, and today’s publications will become training data for tomorrow’s systems. If the scholarly record contains growing volumes of averaged content derived from earlier averaged content, rare cases, unconventional interpretations and genuine novelty may gradually disappear.
Editors are therefore becoming accidental governors of AI. Their decisions will determine which practices are accepted, what authors must disclose and how innovation is balanced against integrity. Yet editors are being handed this responsibility without shared standards, suitable tools or additional resources.
Professor Dobija did not present AI as a tool she personally avoids. At the end of her presentation, she disclosed that the slides had been developed with Claude while stressing that the arguments and judgements were her own. The distinction captured the core of her position: assistance from a tool is acceptable, but it cannot replace human intellectual accountability.
Does AI weaken critical thinking?
The clearest disagreement between the panellists emerged during the discussion of doctoral education. Professor Dikova asked what happens to critical reading, theory building and methodological reasoning when early-career researchers delegate literature reviews, research-question design, method selection and writing to AI.
Professor Morgan-Thomas responded provocatively that academia should reconsider which capabilities the next generation of scholars will actually need. Doctoral education is already misaligned with the academic career system: candidates are trained to produce a thesis but are subsequently rewarded largely for publications. She described an early-career researcher who had made remarkable progress within a few months by working skilfully with AI. The tool itself does not eliminate reflexivity, she argued. Used well, it can help develop it, provided that the researcher is capable of sustaining a demanding dialogue with the model.
Professor Dobija drew a much firmer boundary.
We do need critical thinking, and AI is killing critical thinking, she said. Researchers cannot unconditionally trust an answer produced by a model. They must be able to select what matters, verify the output and add their own judgement.
She did not reject human-AI collaboration, but objected to outsourcing complete intellectual tasks. An automated literature review may omit the newest or paywalled studies and reproduce biases arising from what the model can access. Without prior knowledge, a doctoral researcher may not even recognise what is missing.
Can AI review AI?
The next question concerned automated peer review. In a system overwhelmed by rising submission volumes, could AI move from assisting reviewers to replacing them?
Professor Morgan-Thomas viewed the crisis of the current model as an opportunity for more fundamental redesign. Peer review is already overloaded and frequently fails to perform its developmental function. AI may push the system to a point at which new mechanisms for evaluating and publishing knowledge become unavoidable.
Professor Dobija accepted AI-assisted reviewing on the condition that the tool helps generate questions, organise comments or check elements of a manuscript while final judgement remains human. She opposed fully automated decisions. AI-content detectors generate false positives, and uncritical reliance on them may particularly disadvantage authors who write in highly formal English or use professional language support.
As an editor, she therefore refuses to reject a manuscript automatically on the basis of a detector score alone. Instead, she examines matters such as whether the cited sources exist and whether the argument is coherent. A tool may identify an area requiring attention, but it should not deliver the verdict.
A boundary that must remain visible
An audience question about generating fictitious qualitative interviews showed that AI can do more than support analysis. It can create material that closely resembles empirical evidence. On this point, the panellists agreed. Presenting synthetic statements as interviews conducted with real participants is data fabrication. The tool may be new, but the violation of a fundamental research principle is not.
The discussion did not produce a simple catalogue of permitted and prohibited uses of AI. It did, however, reveal a shared core of accountability. Researchers cannot delegate responsibility for their claims to a tool. Reviewers cannot conceal their own judgement behind an automated assessment. Editors should not replace due verification with the score of an imperfect detector. Universities cannot demand faster publication output without investing in the capabilities, standards and evaluation systems needed to sustain it.
AI does not remove accountability from research. It makes the need to locate it far more precise.
