Abstract
You have likely been exposed to or used Generative AI by now, but AI is more than just text, audio, image and video generation. Traditional AI performs repetitive tasks, predictive AI makes predictions about the future based on past events, and agentic AI makes decisions on your behalf. Each form of AI has unique benefits and risks to every type of user.
In academic publishing, while the effects of agentic AI have yet to be felt, other forms have impacted authors, editors and readers of the Journal of Orthodontics, particularly over the past two years since large language models (LLMs) like ChatGPT, Claude and Bard became publicly available. Authors are using traditional AI and predictive AI for modeling, data and text analysis, and use generative AI in their research, manuscript preparation and formatting. Editors employ predictive AI to find relevant peer reviewers and detect plagiarism. Readers are exploring predictive AI through personalised recommendations from publishers and social media sites and use generative AI summarisation tools to help ingest articles.
The impact on research and information consumption is already significant, and this impact will rapidly increase as AI applications improve, more researchers embrace AI in their work, and more publishers integrate AI-based tools into the peer review process. So, what’s coming? And what are the benefits and risks?
For authors
Pattern recognition is a key aspect of many AI applications, from identifying cancers in medical imaging to self-driving cars’ detection of objects on a road. Any author who has performed statistical or text analyses to identify trends in their datasets or conducted a comprehensive literature search knows that these are intricate and time-consuming tasks. AI applications may take minutes to accomplish these tasks where researchers take days. Consequently, researchers are likely to employ AI routinely for some of these tasks, especially where datasets are large and could take teams of humans too long to process by themselves.
In orthodontics, we may see the emergence of big data analytics, which leverages machine learning and natural language processing to power predictive technologies for clinical use (Allareddy et al., 2019) or facilitating systematic reviews of large and multicentric datasets. However, biases in datasets, such as language (multilingual LLMs are the minority) or treatment norms related to cultural differences, may impact the generalisability of research findings. Orthodontic studies often have smaller sample sizes, raising questions about the ability of LLMs to produce valuable results as they have in fields like genomics, epidemiology or economics. If an LLM’s dataset includes mostly small studies or those from a single centre, overfitting can occur, where an algorithm performs well on its training data but not on new data (Nordblom et al. 2024).
What about using generative AI to write papers? In my experience working with journals, authors are using generative AI due to language barriers, lack of confidence in writing abilities, time constraints, or unclear journal policies. However, with the current state of generative AI’s output, unless you are satisfied with submitting poor quality work, breaking ethical publishing norms and risking detection by the editorial team, it is not advised. Nevertheless, it is likely that software will soon be able to produce unique, high-quality content, raising important ethical questions for the academic community and society at large. For insight into this issue, I recommend exploring the Postplagiarism movement (Eaton, 2024), even if it may be unsettling.
For Editors
I write this editorial as my final act as Editorial Assistant for the Journal of Orthodontics, but it is easy to imagine not just me leaving the role, but the role itself disappearing. Many of the tasks that I did for the Journal and much of the peer review process is likely to be automated in the coming years. Editors already use a variety of these including plagiarism detection software, automated reviewer suggestions, and email generator bots, which can all be enhanced with latest advances in AI. I also have no doubt that some editors will use AI applications to evaluate a paper’s “citeability”, assess authors’ academic or scholarly history, inform decisions on manuscripts, or develop metrics to gauge a paper’s potential to contribute something novel to a particular field. All these tools create ethical concerns, which highlight the need for sufficient human oversight.
For readers
Under Jayne Harrison’s stewardship as Editor-in-Chief, the Journal of Orthodontics is expanding the accessibility of its content. Plain language summaries and video abstracts (Harrison, 2024) enable readers outside the speciality or those preferring alternative mediums to engage with content. Generative AI has a huge role to play in making journals and the internet a more accessible space. If you haven’t already, give Google’s NotebookLM (Google, n.d.) a spin. Share the link to this article or your last published paper and Google will transform it into a podcast where two convincingly human hosts do a remarkable job at providing an audio overview. It won’t be long till we see these kinds of applications integrated into publishers’ websites. But again, there are ethical concerns. While convenient, will researchers rely on these tools and neglect to read the papers themselves?
Now, and later risks
In this guest editorial, I’ve touched on quality, discrimination and transparency concerns, but perhaps the most significant current risk is that people do not trust AI to harness its benefits. The results of Oxford University Press's recent AI survey (OUP, 2024) are fascinating and somewhat surprising in this regard. Early Career Researchers and younger people tend to distrust AI more than older generations of researchers. This may reflect concerns that younger generations have about job displacement, but it could also reflect a phenomenon observed in tech-led information environments. For instance, younger people are more likely to identify AI-generated deepfakes, but they are also more likely to think that a real image is fake. Older people are more likely to believe everything that they see is real. Another notable finding from the survey is the disparity in trust levels between high-income regions and low- and middle-income regions. While most researchers worldwide are optimistic about AI's role in research, those in high-income regions, especially the UK and North America tend to be pessimistic. Improving trust in AI is an important challenge.
As AI enhances our ability to analyse and present data, will it restrict human freedom of thought and creativity? Is that a risk worth taking if AI surpasses human capabilities and leads to groundbreaking discoveries in medicine, materials science, climate solutions and other domains? I foresee a future where AI touches every academic paper to varying degrees, potentially leading to a dividing line in the academic literature where the literature that existed prior to the operationalisation of AI informs the AI-augmented rest.
Between now and whatever future is coming for us, journals have a responsibility to communicate clearly to authors what constitutes acceptable AI use and what does not. When AI has been employed in research, development, or writing, journals should not view this as a red flag. Instead, they must ensure that AI use is transparently disclosed to readers. The Committee on Publication Ethics’ (COPE) statement on the use of AI (COPE, n.d.) echoes this position. Furthermore, journals and universities should organise AI literacy campaigns to educate researchers on ensuring the accuracy of results generated by generative or agentic AI. In the coming age of AI-driven research, fact-checking will be one of researchers’ and editors’ greatest assets.
The opinions expressed in this article are those of the author and do not necessarily reflect the views of CIGI.