Artificial intelligence, large language models, and you

Abstract

Recently, an editorial was published by this journal regarding the role of Artificial Intelligence (AI), and specifically Large Language Models (LLM’s), in vascular surgery research1Exploring the pros and cons of using artificial intelligence in manuscript preparation for scientific journals, Smeds et al.Journal of Vascular Surgery – Cases, Innovations, and Techniques. June 2023; 9101163Google Scholar. Well-trodden contrivance aside, some valid points were raised but, unfortunately, rather superficially addressed in both the AI-generated and Author-written sections. Answers to how this complex and world-altering advancement will affect vascular surgery research, let alone the global repercussions, are beyond even the experts in the field. A deeper analysis, however, is critical. The editorial was focused on the LLM’s impact. To limit the discussion of AI to only how LLM will affect the field of vascular surgery is to miss the wider view of what is coming our way. With the editorial as a starting point, we hope to expound on some of the points made. While some assertions may read as hyperbolic, we encourage the reader to allow for this fact: in a timeframe measured not in decades, but rather months and years, everything is about to change. The section of the editorial written by ChatGPT lays out the usual arguments for and against LLM’s in much the same way a middling student would write an SAT English Writing essay; here is point #1, and here you will find point #2, without elaborating on the relative impact of each. ChatGPT states that LLM’s can ‘potentially save authors a significant amount of time and effort’1Exploring the pros and cons of using artificial intelligence in manuscript preparation for scientific journals, Smeds et al.Journal of Vascular Surgery – Cases, Innovations, and Techniques. June 2023; 9101163Google Scholar before moving swiftly on. This is, however, a gross understatement of the potential impact that should not be left as a simple one-liner. A more accurate statement would be that Large Language Models, in combination with data analysis AI, will save authors time and effort by completely eliminating humans from the academic research process. When an engineer properly turns their gaze on the medical research world, the days are numbered for the traditional Academic Surgeon. Say goodbye to the medical student who is so efficient at literature review; goodbye to the statistician on staff to assist in data analysis; goodbye to the Lead Author editing and scrupulously wording each statement. Literature review has for a long time been an online activity, making it perfectly suited to automation. LitReview 2.0 will be able to analyse the research articles it may cite to find those with adequate sample size, randomisation, and sufficient power for statistical significance. It will consider the methodology, exclusion criteria, and limitations. And it will do it faster and better than Jack or Jill MS3. It will be able to poke holes in Discussion sections based on the 100,000 articles it ‘read’ within a few seconds and come to conclusions based on the best available data. Take the cumulative literature reviewing and literature consuming experience of every vascular surgeon currently practicing and compress it into a single application. Comprehensive literature review at the click of a button. Data analysis has long been under the purview of statistical computer applications. With the growing number of large, multinational databases with increasingly granular detail, crunching the numbers on 3,000 people across the world who have undergone a carotid endarterectomy will take a couple of keystrokes. ‘Does reversing with protamine decrease haematoma formation after carotid endarterectomy?’ ‘Here are the demographics, risk factors, Cox regression, and Kaplan-Meier curves, Doctor. And may I suggest we also investigate if the haematoma was symptomatic or required surgery?’ ‘Why thank you, Hal, that would be great.’ Dearest to the hearts of many of the readers of this journal is how the all-important Lead Author is affected by all of this. The insight and experience of the Lead Author to come up with the original research question and determine which salient details should be most prominently discussed must be indispensable. The other bits can be automated, but how will we know that this research was reviewed by a trusted source – the Lead Author. We would argue, however, that insight and experience are merely data points collected and organised in the wetware that is your brain. Research questions often arise from anecdotal experience: ‘I’ve noticed that when I do X, sometimes Y happens.’ Then the research army gets to work finding evidence to support or refute the anecdotal observation. Alternatively, an author reads an article and has further questions about a subgroup or another outcome they are interested in. Both of these ‘creative’ processes are artificially reproducible. As data acquisition increases, each moment of a patient’s stay will be recorded and stored. Currently, you can monitor your patient’s vitals from anywhere in the world with a secure network. Put these data points in a server that can be accessed for research purposes, and AI can be doggedly running analyses to find connections between every X variable and every Y variable that it is possible to record. There will no longer be a need for (1) an anecdotal observation, to spark (2) a research question, that leads to (3) analysis of patient data to test the hypothesis. It will be all 3’s, all the time. Certainly, at first, the Lead Author will remain the only member of the research team left standing. Consider, however, an analogy. We do not presume to speak for others, but when we do our multivariable regression analysis or propensity score matching, we do not crunch some of the numbers ourselves on a giant chalk board to verify the algorithm is correct. We trust in the algorithm; it has been verified. Given time, as more and more journal articles are produced with less and less input from various members of the research team, the confidence in the algorithm will grow and ultimately, will be beyond question. There will be periodic checks and random sampling to ensure nothing is amiss (done primarily by yet another algorithm). Much like the Excel or SPSS algorithms, however, these checks will be done by someone else, somewhere else, and future generations of surgeons will take little notice. Scepticism about such fantastical claims is understandable, especially given the sparsity of citations, but any observer of the recent trends in technological advancement will not find much at fault with this version of the future that is synergistic with the exponential increase in computing power. Await the rise of widely available quantum computing, and when it arrives, sit back as your phone alerts you with a simple message from the JVS App: ‘New EVAR indication is 4.5cm AAA.’ If robots haven’t already taken endovascular procedures out of the hands of surgeons, you will have received your marching orders. By no means do we claim to have addressed all aspects of this vast topic. And, of course, there will be a learning curve for the AI in the meantime; plenty of glitches and examples of failure that will be held up by Luddites as a casus belli against the whole idea. One prominent argument is that the current generation of AI is ‘only as good as the data they are trained on’1Exploring the pros and cons of using artificial intelligence in manuscript preparation for scientific journals, Smeds et al.Journal of Vascular Surgery – Cases, Innovations, and Techniques. June 2023; 9101163Google Scholar. While this is valid, the same can be said of academics themselves, and such is why peer review is in place. That is, until it is no longer required (see above.) Another is regarding ‘who gets the credit’1Exploring the pros and cons of using artificial intelligence in manuscript preparation for scientific journals, Smeds et al.Journal of Vascular Surgery – Cases, Innovations, and Techniques. June 2023; 9101163Google Scholar. This is where we believe the crux of the current AI Panic lies in academia. Existential threats to certain segments of society provoke a reactive (apolitical) conservatism that can stymie a nascent revolution. By all means, AI-generated research is an existential threat to every academic field in existence. In this vein, we ask that the current up-and-coming generation of Academic Vascular Surgeons investigate their motivations and ensure they are not approaching academia as a ‘surrogate activity’2F. C. The Unabomber Manifesto : Industrial Society and Its Future. Berkeley, CA : Jolly Roger Press,[publisher not identified], 1995Google Scholar. The goal is to improve and lengthen patient’s lives, not your resumé. During the advent of endovascular interventions, we staked our claim at the forefront of a new paradigm. This should be no different.

Ähnliche Arbeiten