Over a year ago, I posted a rather lengthy rant concerning the contemporary pestilence in academia that is the use of generative artificial intelligence, which I will abbreviate to GAI here, in the process of writing academic manuscripts and more egregiously, the creation of figures and diagrams. In that essay, I made my stance on the issue rather clear — that GAI has no place in the writing of and production of academic content, especially in the main texts of manuscripts and scientific visualisations. Today, I want to revisit the problem, and discuss how the problem has developed over the past year or so.

I want to start off by talking about the bright side of things. Fortunately, there has been an increasing awareness shed upon this issue, from creators like me who discussed at length about the problem and its implications, to academics and enthusiasts who openly voiced their disdain for GAI in academia. The current body of academia seem to be overwhelmingly against the use of GAI in the creation and production of academic publications, though the community is split in how this problem should be mitigated or curbed. Nevertheless, I believe that awareness of such an issue is a start to dealing with this problem.

Furthermore, there are more accessible resources that identify suspected or alleged GAI use in academic papers and publications, some of which I have found through Retraction Watch. One such example is Academ-AI, created by Alex Glynn from the University of Louisville in Kentucky. On this website, journal articles and conference papers or proceedings suspected to have GAI involvement are documented, and are identified through signatures left baked into the manuscripts (and/or figures). Additionally, he has also provided further advice on what to look out for in contributions made by GAI or even chatbots. This website has been a great resource, and I recommend checking it out. Interestingly, and perhaps ironically, the logo used is a vector image based on what Google’s Gemini generated.

But of course, this is not all sunshine and rainbows. We could discuss the increasing awareness made on this brand of academic dishonesty for hours, but it is also important to examine how the magnitude or prevalence of this problem has changed over the year. As an independent (and informal) search, I selected PubMed as the main literature database to search for GAI signatures. For the uninitiated, PubMed is an openly-accessible database of publications the fields of life sciences and biomedical sciences. On its website, it consists of more than 38 million citations for such publications, making it one of the largest academic databases for these disciplines.

I recall a video essay made by a YouTube channel called Upper Echelon, which also explained his search strategy on PubMed. I decided to adapt this strategy with extra terms and signatures identified and picked up by Academ-AI, and conducted my search for publications, with a particular interest in trends over the last 10 years. As a visual aid, PubMed provides a graph visualising trends in the number of publications containing the search term. It must be noted that academic interest in life sciences and biomedical sciences has grown over the years, and could be thought of as multicollinear or perhaps confounding with the time trend of GAI involvement in academic publications. After all, with the increase in academic interest and activity, so too would academic dishonesty of this nature. Nevertheless, when we search for publications containing only the GAI signatures as strings, we should expect that the prevalence of such articles be low but gradually increasing in the early 21st century, or even in the 2010s, with a rapid increase in the 2020s as GAI models enter popular use. Without GAI, we should expect these trends to gradually increase, without this spike in 2023 to 2024.

The first string of note is the term ‘meticulous’. As with other excessive words such as ‘commendable’, ‘meticulous’ is amongst the most well-known signatures left in text by ChatGPT. This is also one of the search terms covered by Upper Echelon, and after replicating his little search, I was presented with this:

PubMed publications containing this very term spiked in 2024, though an increase actually started in 2023, likely towards the tail end of the year as ChatGPT grew in popularity and usage. As of 26 March 2025, the number of articles containing the word ‘meticulous’ has already exceeded that in publications prior to 2020. The way I framed these screenshots does include the titles of some publications with that search term, but I did not verify its full text for GAI signatures as I wanted to include the main search filter used. Thus, there is a high possibility that these articles are written by humans which just happen to include words that tend to be used by GAI.

Moving down Upper Echelon’s list of GAI signatures, there are also words like ‘delve’, ‘realm’, ‘notably’, and ‘notable’, all of which have demonstrated a spike in 2024. Going back to Upper Echelon’s video, his brief analyses mainly looked at 2023 alone. This has shown how rapidly the use of GAI has taken over scholarly works, and this is in PubMed alone. It should be noted that the use of these terms have generally been increasing in the years leading up to today, owing to the growing scientific and academic body. However, comparing the rates of increase in the years before ChatGPT’s public release in late-2022 with the spike we see, this spike does seem to be an outlier that can be explained by the use of GAI in manuscript ‘writing’ and scientific figures. There are some genuine articles written by humans that have showed up as false positives, and there are some resources and advice to help human authors to not be falsely identified as GAI, such as the avoidance of textual GAI signatures including the ones shown in this brief PubMed search.

It is almost certain that the infiltration of GAI in scholastic works is getting even more entrenched, and it is an increasingly growing issue that journals and conferences have not been able to catch up on. Sure, some publishers have outlined policies on how the review process should take into account the use of GAI in content generation, but how stringent is this? Are there any sanctions or penalties incurred should reviewers fail to check for GAI involvement? Given this increase in publications containing GAI hallmarks, it does seem that publishers, journals, and reviewers are still not doing enough to weed out these publications.

Nevertheless, the number of retracted studies continues to rise, with the total number of retraction entries nearly crossing the 55000 mark last year. Observing these current trends in the use of GAI in manuscripts and figures, we can only expect this number to balloon this year. This is also not mentioning the rise of competing large language models in the GAI ecosystem, such as Google’s Gemini, Twitter’s Grok, and Deepseek. Each of these models are bound to have their own identifiable hallmarks and modus operandi, and so identification of involvement of other LLMs in generating scholarly content might be more difficult.

For a silver lining, this phenomenon has generated academic interest in understanding the ethics, morality, and rationality of GAI’s involvement in academia. In the past couple of years, several articles have been published covering the propensity of individuals to use GAI in plagiarism or academic dishonesty. Zhang, Amos, and Pentina’s article, for example, suggested that people experiencing ‘moral disengagement’ (the belief that some ethical standards do not apply to them) were more likely to commit GAI-based plagiarism, with other factors being the consideration of benefits and possible sanctions.

Strzelecki’s 2025 publication highlighted the methods used to identify the hallmarks of GAI in scholarly works. Compared to Academ-AI, this publication goes into greater empirical detail in the distribution of GAI use by discipline, and even by journal. In his study, medicine, computer science, and engineering are the disciplines in the sample that contained the most number of publications with GAI hallmarks. He also noted in his discussion the existing methods of dealing with this issue, including the use of OpenAI’s tool to identify ChatGPT-generated content. These methods are quite limited and unreliable, leading to many instances of false positives and negatives. He also touched on the need for the entire publication process to include the identification of GAI-generated content, something which I heavily touched upon in my 2024 essay. I highly recommend checking this article out, it is open-access, and I think it is an important read in this current climate.

Now, in 2025, I am writing about this as a postgraduate student, and have reflected upon what I have written in last year’s entry. I remain steadfast in my staunch opposition against the use of GAI in the production of scholarly works, manuscripts, and publications. I have mentioned this a couple of times prior, and I will mention this again, that GAI has no role in academia, and its use should not be encouraged or even permitted in its institutions. I have encountered several cases of AI-generated images used in presentations, seminars, coursework, and even dissertation cover images, and I am sick of them. I have also voiced my opposition against the use of GAI my research projects to my supervisors, with which they have agreed. But honestly, I wish institutions and universities did more to regulate and enforce, or better yet, prohibit its use in research and coursework. We need these regulations and enforcement, and we needed them yesterday. The infiltration of GAI has exacerbated the undermining of the credibility of science, research, and academia, and we cannot just sit idly by and do the ‘bare minimum’; we will only be shooting ourselves in the foot.