How One Oxford Student Used AI to Commit Academic Misconduct at Industrial Scale
And what it means for the quality of peer-reviewed conferences
The world of research is competitive. PhD students need to publish to graduate, professors to get grants, industrial labs — for prestige and acquiring top talents. It is not surprising that many got tempted by using LLMs to generate research papers. Emily Bender and colleagues coined the term "stochastic parrot" and parroting research is the plague currently rampaging in the world of peer-reviewed conferences. The submissions are filled with meaningless strings of words that might sound like research and be grammatically correct but essentially have no meaning.
Noam Chomsky used the phrase "Colorless green ideas sleep furiously" to illustrate that grammatically correct sentences can be semantically nonsensical. In Syntactic Structures (1957), he argued that probabilistic models are inadequate to capture linguistic competence: both the grammatical but meaningless sentence and its scrambled, ungrammatical counterpart would receive identical near-zero probability scores from any statistical model, yet any competent speaker knows one is well-formed and the other is not.
The colourless green ideas of the modern conference paper submission sound like this: "Quantum-Aware Generative AI for Materials Discovery: A Framework for Robust Exploration Beyond DFT Biases." Or like this: "Neuro-Symbolic Hypothesis Engine: A Unified Architecture for Autonomous Scientific Hypothesis Generation."
Each of these scientific buzzwords makes sense on its own. They are combined coherently and grammatically correctly. But together they say nothing.
Those are the titles of the papers by Mahule Roy.
Just a couple of weeks ago, if you were to google this name, you would find many papers published by this individual. Their Google Scholar account (now deleted) looked like an account of an established researcher: dozens of papers, co-authored with well-known respected scientists, published in various venues, journals, or as pre-prints on arXiv.
This aligns with the profile of Mahule too: a researcher affiliated with Kellogg College of Oxford University. Kellogg is a real, legitimate Oxford college. Nothing about that affiliation would raise an eyebrow.
At first sight, nothing looks suspicious here.
Ivan Habernal, Professor of Trustworthy Human Language Technologies at Ruhr University Bochum, would normally not have given a second glance to this Google Scholar profile. Yet Prof. Habernal is also an organiser of the PrivateNLP Workshop, co-located with ACL, the Association for Computational Linguistics - the scientific and professional organisation that conducts top conferences in the area of AI and NLP. This is what Prof. Habernal wrote on his LinkedIn:
Mahule Roy and co-author Subhas Roy, assumingly Mahule's alter ego, who listed their affiliation as TATA Consumer Products Limited [1] , are very well known among the organisers of workshops co-located with top conferences such as AAAI, ICLR, EACL, and ACL.
Mahule Roy and Subhas Roy's claim to fame is nothing but exceptional productivity and the ability to publish on a wide variety of topics: from African NLP to polymer informatics.
Science Runs on Volunteers
In the research world, publishing is very competitive, publishing is needed to get grants, scholarships, tenures and in general improve the prestige of the group. As NLP and AI became hot topics - ACL conferences (top NLP conferences) have seen a raise in submissions. The acceptance rate is low and many prominent research labs with large budgets such as Google or Meta submit papers to the conferences. Last year ACL has been dominated by researchers from China and the best paper award was given to DeepSeek. In this situation it is not easy to publish in top conferences particularly if your focus of interest lies outside of the LLMs - for example on low resources languages, trustworthy NLP, historical languages and more.
Workshops co-located with top conferences are a way to publish work that might interest a more niche audience and to present it in smaller circles. Those workshops are more inclusive, the acceptance rate is significantly higher, and sometimes it is even sufficient to submit an abstract rather than a full paper.
The workshops are organised by volunteers who do this work without additional compensation. These are normally established researchers whose sphere of interest overlaps with the topic of the workshop. For example, Prof. Ivan Habernal, being a well-known researcher in trustworthy NLP and data privacy, has been organising the PrivateNLP workshop for many years.
Submissions to the workshops are frequently peer reviewed in the same manner as main conference submissions. The workshop chair also meta-reviews the papers, makes the final decisions, coordinates the logistics, maintains the website of the workshop and handles all communication with authors and reviewers — on a purely voluntary basis.
The Workshop Plague: Wasting Researchers' Time
Mahule Roy has submitted and published across a wide variety of workshops covering the most colourful mix of topics: remote sensing, AI for agriculture, long video creation, NLP for African languages[2], and fact verification (now deleted). These are just a few papers accepted within a single year. Many more were published as pre-prints or submitted to arXiv. Mahule Roy's Google Scholar boasted an incredible productivity:
For an untrained eye this profile looks great and convincing. Many of those papers were cited by… Mahule Roy herself.
Mahule Roy is well known among workshop organisers. She submitted to workshops at almost every major AI conference. When Prof. Ivan Habernal posted about her on LinkedIn, many other researchers spoke up.
And I am convinced that after this article is published, many other researchers will remember these so-called authors.
Yet the problem is not as obvious as one might think. It is hard to pinpoint exactly what Mahule Roy is doing wrong. Most workshops do not prohibit the use of AI, and how could they?
First, there is no bulletproof method to prove that something was purely generated by AI. Second, using AI for correction and paraphrasing is not uncommon. I have seen best paper awards at top conferences given to researchers who barely speak English. In no way do their language skills diminish their scientific contribution, and if they used a model to correct their English, the result can look like an AI-generated paper. Still, the problem is enormous for workshop organisers. As Prof. Ivan Habernal told AI Realist:
This behavior is particularly outrageous, as the workshop organization and reviewing are done on a purely voluntary basis.
In other words, once such a paper is submitted, many people will be drawn into reviewing it. It will normally be sent to three reviewers, other researchers who will anonymously peer review it, read it, and write their assessment. It is very well possible that they will be the first people ever to read this paper, given that it is highly doubtful Mahule Roy read her own "publications." Three researchers will, on a purely voluntary basis, have to spend many hours reading AI slop. A paper from 2013 estimates that it takes on average 11,5 hours to review a paper:
Even if the current review process can be, to some extent, optimised through the use of AI tools, it still takes a significant amount of time.
The reviews will then be sent back to the workshop chairs, and a meta-reviewer if applicable, or the workshop organisers, will need to process them and make a decision. Thus, Mahule Roy wasted days if not months of work from people who could have spent their time in a far more valuable way.
AI Realist is an independent, anti-hype, vendor-agnostic publication. We also offer workshops, seminars, and individual advisory services. The work is fully funded by subscribers and the trainings and seminars that AI Realist offers. To support us and get insights into AI without hype or overpromising, become a paid subscriber.
Name Theft And Discrediting Researchers
Where Mahule Roy undoubtedly committed serious misconduct goes beyond simply generating papers: she was adding well-known researchers as co-authors without their knowledge.
AI Realist reached out to those who allegedly co-authored papers with Mahule Roy. One of them is Guillaume Lambard, a Senior Researcher at the National Institute for Materials Science in Tsukuba, Japan. The nonsensically titled paper mentioned at the beginning of this article lists Dr. Lambard as a co-author.
Dr. Lambard is one of the world's leading researchers in materials science, with over 4,000 citations on Google Scholar. One can only imagine how outraged he was when he discovered that he had "co-authored" this nonsense:
“Indeed, my name has been associated with it without my knowledge. — Dr. Lampard told to AI Realist — I have discovered this problem a few weeks ago while submitting a manuscript on ArXiv. The fraudulent paper has since been removed from the ArXiv thanks to the efficient support from the ArXiv's staff. I firmly condemned any kind of fraudulent behavior, the present one included.”
Dr. Lampard is not the only one whose name Mahule Roy fraudulently added to her work. Her Github account, dreamboat26, which she now also deleted or made private boast co-operation with top researchers such as Prof. Dr. Sarfaraz K. Niazi, a world famous expert in pharmaceutical science. Mahule claimed to have done a short-term project under his guidance:
Prof. Niazi unsurprisingly has never heard of her:
“He[3] claims to be a remote employee of Purdue University and to have a degree from Oxford; I have never heard of him or of him contacting me. I have looked at his publications, and they all appear fraudulent to me. This must be fully investigated, and you can quote me denying any association with him.” — he told to AI Realist.
An internet search found another paper that was withdrawn because one of the co-authors had not been informed. Considering that there are only three authors, two of whom are assumedly Mahule herself, we might assume that the unaware "author" was supposed to be Omar Abudayyeh, Assistant Professor, Harvard Medical School:
Fictional Affiliations and Discrediting Institutions
At this point in the investigation, it was still unclear to me who this person is and why they are doing it. The Harvard Medical School affiliation was one that appeared on multiple papers, and it is also the affiliation of one of the people she added as a co-author without their knowledge.
She also listed it on presentations at various workshops:
AI Realist reached out to Harvard Medical School for clarification and got a clear response:
“We have no record of a Mahule Roy at Harvard Medical School, either as a student, employee, or holding an academic appointment.”
Under further investigation, it turned out that she “affiliated” with a whole lot of prestigious universities. Her github listed Carnegie Mellon University, now deleted LinkedIn had Purdue University.
Her website[4], now also taken down, but still available as a snapshot from 15th of July, 2025, lists a whole range of affiliations:
Yet AI Realist found out who Mahule Roy really is.
Who Is Mahule Roy
So far the story seemed quite trivial: an anonymous person on the internet had created a number of fake profiles and was spamming workshop conferences. Then the story took a completely unexpected turn.
Mahule Roy is a real person. She is a master student at Kellogg College, Oxford University.
The shocking twist was that not all her papers are fraudulent. In the pile of AI slop, there was one paper that stood out: “Geometry of Self-Supervised Embeddings Reflects Bone Tissue Structure”, submitted to a workshop on Medical Imaging with Deep Learning. The paper was co-authored with Prof. Geert Litjens and Dr. Gerry L. Koons from Radboud University Medical Center. Even at first glance the paper looked like it had actually been written by a researcher: paragraphs were polished, diagrams made sense, tables were aligned. AI Realist reached out to the researchers and was genuinely surprised:
Mahule Roy reached out to us at the start of the year with an interest in doing a remote internship with our group, of which the paper you speak of is a result. [….]
That said, for our work, this was based on real data from my lab and was also inspected by a board-certified pathologist. As such, the data in the paper in question was real and vetted by both Dr. Koons and me and our names were used with our consent. That of course does not preclude misconduct in other work, but that I cannot speak to. - said Dr. Litjens
The paper was rejected from the workshop and now Dr. Koon and Dr. Litjens fully retracted the submission.
Another paper that did not produce an impression of AI slop was “Device-Robust Spectral Grading and Origin Detection from UV-Vis-NIR Images: Towards Practical Gemstone Quality Assessment“. It was co-authored with the Institute of Applied Artificial Intelligence and Robotics in Tennessee, USA, and the Applied Artificial Intelligence Initiative at Deakin University, Australia. As one of the co-authors, Dr. Srikanth Thudumu told AI Realist that Mahule had applied for a summer internship and was selected from among other candidates. She is described as engaged and proactive during her time there, with this paper being the result.
Finally, the last paper that did not appear to be AI slop was Autonomous Multi-Agent AI for High-Throughput Polymer Informatics: From Property Prediction to Generative Design Across Synthetic and Bio-Polymers. Prof. Zhang confirmed that the paper is not fraudulent:
The arXiv preprint you referenced (2602.00103) is a legitimate research collaboration between my PhD student, Adib Bazgir, and Mahule Roy, whom he met at a Stanford hackathon. I supervised the work as corresponding author.
The picture was becoming clearer and clearer. Mahule Roy is not an internet troll. She is a student who is genuinely eager to do research. Her website also listed a number of achievements common for enthusiastic and high-performing students, such as participation in hackathons and coding challenges.
Among all the affiliations Mahule Roy listed, one was the most consistent: Kellogg College, Oxford University. Several of her papers also cited an email address ending with “@kellogg.ox.ac.uk”.
AI Realist managed to reach Mahule Roy on her email for an official statement:
“I am extremely sorry for my behavior and I want to truly apologize. I did use AI for drafting and few early cases to use AI for citations and since then I haven’t used AI for anything more than that. I have since then withdrawn all my work and over mail rectified the mistakes I made by using Dr. Omar’s name without his consent.” - she wrote in her reply to AI Realist
In her initial statement, Mahule stated that she had used AI for drafting her numerous papers early in her work and had since stopped doing so. I will leave it to the readers to judge the truthfulness of this statement.
GenAI for Scientific Writing: Clear Guidelines, Unclear Enforcement
I approached Kellogg College for an official statement on the situation. It has been confirmed by both Mahule and Kellogg College that she is indeed enrolled at Kellogg College as a master’s student.
The initial reply and subsequent replies focused on the validity of the claims that the papers are AI-generated and on whether the workshops’ policies prohibit such use. That is where the argumentation genuinely gets complex. Arguments such as “the paper is complete nonsense” or “she produced dozens of papers within a few months” are not, in themselves, strong arguments.
As Jeremy Gibbons, Professor of Computing and Senior Tutor at Kellogg College, told AI Realist in correspondence:
“You don’t need GenAI to produce nonsense.”
This is something that is truly challenging for the scientific community. Prof. Gibbons statement is correct: you do not need GenAI to produce nonsense. But you do need GenAI to produce nonsense at scale. Bad submissions existed long before ChatGPT. Back in 2005, three MIT students created SCIgen, a program that generated random computer science papers using context-free grammar, and got one accepted at a conference. The tool got misused and it eventually led to the retraction of 122 papers.
Creating a bad paper of scientifically sounding nonsense is not new but it used to take time and effort to produce a paper that is good enough to be accepted. It is not the case any more. The problem we are dealing with today is quantity.
ACL deals with AI-generated articles in a clear and straightforward way: any use of AI must be disclosed, and the information must be validated by the authors. Even if the generative tools came up with the ideas:
New ideas. This covers when generative model output reads to the authors as new research ideas that would deserve co-authorship or acknowledgment from a human colleague (e.g., topics to discuss, framing of the problem), which the authors then develop themselves. As with all new ideas, the authors should conduct a literature search to determine relevant prior work and cite to ensure proper credit. The authors should disclose if models were used in this manner.
Unapologetically, ACL does not tolerate any absurdities like adding Claude as a co-author:
New ideas + new text: ACL does not consider a generative model to be an entity that can fulfill the requirements of co-authorship.
The enforcement of these guidelines, however, is a difficult challenge. AI detection tools are unreliable: even if a paper is flagged as AI-generated, a manual review by the programme committee is still needed. Even such manual reviews cost time and effort, and cannot compare with the scale and speed at which AI slop is produced by "authors" such as Mahule Roy.
That these guidelines are hard to enforce even for ACL organisers, who are among the world's top researchers and experts in NLP and large language models, is confirmed by their recent announcement:
The official statement published by the ACL programme chairs explains that their automated system checked the citation consistency of accepted papers and flagged over 100 with hallucinated citations. They declare:
An author who fabricates a citation commits a serious breach of ethics, and using an automated system as a proxy to generate such citations is equally unacceptable.
The papers accepted to ACL are not comparable to the AI slop published by Mahule Roy. The rejected papers are not fraudulent: they are genuinely high-quality, innovative work that required months of research. The AI-generated citations are an unfortunate oversight on the part of the authors. If you have ever written a publication, you know how related work research relies on Google searches and, increasingly, on LLMs. It is not uncommon to delegate this to junior team members, such as students or interns. The rejection of these papers is undoubtedly a very unfortunate event for those who had already booked flights and celebrated their acceptance. Yet strict guidelines and their enforcement are absolutely necessary for maintaining the integrity of scientific work.
My personal opinion, though, is that protecting scientific integrity and the quality of research should not be solely the job of conference organisers. Research institutions should introduce strict guidelines on the use of generative tools and the submission of AI slop under their affiliation.
The Response of Kellogg College
Kellogg College, Oxford University, did not provide any statement on the behaviour of Mahule Roy or reference to any guidelines and measures that can be taken to prevent or allow, in case they consider it acceptable, this practice under their affiliation.
I cannot comment on the content of the submissions.
Jeremy Gibbons told to AI Realist in addition to the comment quoted above that nonsense is not always AI generated and it is unclear if workshop guidelines prohibit it.
Kellogg College explicitly refused to comment on the quality of the submissions, nor indicated how they evaluate this incident: unacceptable, acceptable, not their business? These questions remain open for the wider public.
On the question of whether it is acceptable to assign authorship without consent, Kellogg College clearly stated:
Adding the name of some unrelated person to your paper is obviously completely unacceptable behaviour
My personal opinion is that solely questioning whether conference organisers added appropriate guidelines, without having a clear answer on what the institution's position is on the submission of AI-generated texts, is insufficient.
Mahule Roy's submissions are ridden with hallucinated citations, references to non-existing benchmarks, serious methodological flaws, misinterpretations, and outright nonsense. Of course, all of this can happen without AI, but the damage is not about quality: it is about scale, and it should be treated differently as one would treat a simple mistake by a struggling student. Submitting one poorly written paper is a forgivable offence. Submitting dozens should not leave the primary concern as merely whether the GenAI guidelines of all those workshops were technically violated.
Yet I will leave it as an open question for Kellogg College and other institutions to answer.
In conclusion, I want to give the floor to Mahule Roy, who made an official statement regarding her behaviour.
I asked her to make a statement and I will publish it in full:
Dear Maria,
Thank you for giving me the opportunity to respond.
I sincerely apologize for the mistakes and poor judgment involved in some of my recent academic activities. I take full responsibility for the issues that arose, including the inappropriate use of another researcher’s name without proper consent. I understand why this has raised serious concerns.
Since recognizing these problems, I have taken steps to correct them, including withdrawing the affected work and contacting relevant individuals to address the situation directly. I also understand the importance of maintaining proper academic standards and integrity going forward.
I want to emphasize that I am taking this matter seriously and learning from it. I am committed to ensuring that such mistakes are never repeated.
I appreciate the opportunity to provide my perspective and would respectfully ask that my corrective actions and willingness to take responsibility also be taken into consideration.
Best regards,
A Systemic Problem Needs a Unified Front
On my part, I would like to add that Mahule created the impression of a hard-working, genuinely engaged student. What went wrong here is unclear: whether, as a junior researcher, she did not fully understand how unacceptable her behaviour is; whether she was confused by Kellogg College’s policies on the subject; or whether she was fully aware that she was committing serious misconduct. We will never know. What we should pay attention to is that Mahule Roy should not be viewed as a scapegoat. Workshops and conferences receive hundreds of AI-generated submissions. The issue is systematic: being blacklisted by a conference is unpleasant, but what other consequences will these “authors” face?
AI-generated paper submissions are rarely fully anonymous, for one simple reason: they are there to create an illusion of scientific credibility for a particular individual, or to advance a scientific career, for example by fulfilling the required number of publications to complete a PhD. You cannot monetise a Google Scholar page, and having many citations in ACL workshops will not help anyone make millions on TikTok. Those submissions can be traced to companies, institutions, and individuals. And those submissions do create reputational risk for everyone involved.
I believe the war on AI slop in research must have a unified front, and I want to ask my readers affiliated with research institutions or industrial research labs: do you have clear guidelines on what is acceptable and what is not in terms of the use of generative tools when producing research work? If not, perhaps it is time to make one.
[1] AI Realist has contacted TATA Consumer Products Limited for an official statement and comments on whether these individuals are affiliated with them but did not hear back.
[2] The organisers of the AfricanNLP Workshop were informed and provided the following statement to AI Realist:
The submission that you have informed us about is a non-archival abstract, presented for discussion at the workshop. It is not included in the proceedings. While we investigate, we have removed it from the workshop website.
— Prof. Muhammad Shamsuddeen
[3] At the time of communication, the real identity and gender of Mahule Roy was unknown to AI Realist
[4] The website listed her as a graduate of National Institute of Technology Karnataka, Surathkal. Bachelor of Technology; GPA: 8.48/10. Specialization in Artificial Engineering; GPA: 9.0/10 and a former intern of TATA Consumer Products Limited. Neither one responded to the inquiries of AI Realist.
Thank you for this article. It is putting to a problem that science is facing and we should take it very seriously.
I think you the readers will benefit from initiatives such as: https://arxiv.org/pdf/2605.08586
https://promise-check.vercel.app
https://medium.com/@mamadouk/computational-experiments-nonrepudiation-8cc53cdf206a
Thank you for sharing this article. Very interesting, and the topic is of importance.