Studies have increasingly shown the widespread use of generative AI in research publications. Faced with the consequent uptick in the number of publications, Simone Ragavooloo argues that editors and reviewers should embrace AI tools to undertake the heavy lifting of statistical and methodological review and to allow them to focus on areas that require human expertise.

Enjoying this post? You can read all our featured work on Peer Review here. Want even more? Then sign up to our newsletter and receive alerts for all our latest posts here. 

Artificial Intelligence is transforming science and science publishing must keep pace with this change. The World Economic Forum’s Top 10 Emerging Technologies of 2024 report highlights the billions of funding being ploughed into AI in scientific discovery alone.

AI is now already widely applied in research, from discovering new families of antibiotics to studying myriad social and cultural phenomena. The United States’ President’s Council of Advisors on Science and Technology (PCAST) has stated “AI has the potential to transform every scientific discipline and many aspects of the way we conduct science.” AI’s transformative potential lies not only in how we do research, but in how much scientific research we produce, as recognised in the OECD’s 2023 Artificial Intelligence in Science report: “raising the productivity of research could be the most economically and socially valuable of all the uses of AI”.

Publishers must now adapt and innovate just as they did during the shift from print to digital at the end of the 20^th century. However, peer review presents a challenge to these visions. 100 million hours were estimated to be spent on peer review in 2020, a figure that could rise exponentially if reviewers are not supported. Given that the current system is already viewed by some as working at capacity, Lisa Messeri and M J. Crockett have argued an AI-enabled ‘science-at-volume’ could lead to the ‘illusion of understanding’, whereby a significant escalation in scientific productivity and output is not matched by human insight and judgement.

100 million hours were estimated to be spent on peer review in 2020, a figure that could rise exponentially if reviewers are not supported

One answer is to meet like with like. To free up human reviewer expertise, we need to develop and trust more in AI-enabled peer review and research integrity tools to do the heavy lifting. Preventing fraudulent and plain, old ‘bad’ science entering the peer review process is the first step. The situation in research integrity is analogous here with cyber security’s and the finance sector’s use of AI to fight fire with fire. These applications highlight AI’s capability to process vast amounts of data and identify anomalies at a rate that cannot be matched by human detection. Indeed, AI-tools in research integrity already exist, Frontiers’ AIRA came online as early as 2018 and is now joined by an array of AI-tools tackling various aspects of research fraud. The International Association of Scientific, Technical and Medical Publishers (STM) recently created the STM Integrity Hub to aggregate and harness such technological innovations across research publishers.

Positive as these steps are by publishers to protect research integrity, the greatest challenge facing publishers in relation to AI is not malign practices by the few, but the positive adoption of AI tools to advance and expediate research by the many.

So, what is to be done? Publishers need to get past initial limitations with AI and early large language models (including the data available to train these LLMs) and realise the potential of AI-enabled peer review. Open data is an early example of this trend. A core tenet of the open science movement, open data enables AI-in-scientific-discovery to make the connection between interoperable data produced by different research teams.  As AI-enabled scientific data becomes bigger and more complex, the key reviewer task of spotting methodological and statistical errors in submissions becomes more demanding. A situation that is often exacerbated by the lack of advanced statistical training and expertise for some researchers. The combination of AI with open science/open data therefore has the potential to increase scientific discovery and innovation, but it also creates more complex combinations of data and more risk of flaws being introduced to datasets.

To give a real-world example, a-leading science team produced original data with good intent, using machine-learning to identify microbiomes associated with cancer. Post-publication peer scrutiny identified problems with the data and pointed to the “flurry” of subsequent studies that used this data believing it to be sound. Retractions and related investigations followed.  From a publisher and peer review perspective, the question is how the data could have been prevented from entering the scientific record? In this respect, we are all still in the transitional period, where researchers and publishers continue to learn from such incidents and adapt research methodologies and peer-review protocols as the use of AI and LLMs in research becomes widespread.

As more applications of AI in scientific research come online, it is neither desirable nor feasible to rely on a small pool of statistical reviewers to do the heavy lifting with evermore complex data produced at a faster rate

Publishers have the scale and technological expertise to experiment and develop tools in this space. As more applications of AI in scientific research come online, it is neither desirable nor feasible to rely on a small pool of statistical reviewers to do the heavy lifting with evermore complex data produced at a faster rate.  Publishers’ tools should assist both authors and reviewers by automatically and accurately detecting statistical errors or anomalies, suggesting appropriate statistical methods, and providing a preliminary analysis of researchers’ data. If we can achieve this, then even with big, AI-generated datasets, the review process will remain robust but streamlined, freeing up human reviewers to focus on other critical aspects of the manuscript.

Two things are clear. First, peer review cannot be sustained in its current state as AI increases science output. Second, as the volume of research grows, collaboration, as well as innovation, is essential to protect scientific discourse and the integrity of the scientific record. What does cross-publishing, cross-research cooperation look like, from the lab bench to publisher’s page and how do we drive it forward? Can we develop AI tools advanced enough to catch faulty big data before it enters the scientific record? How would a publisher-wide alert system (not dissimilar to cyber security alerts) work, to share intelligence that stops flawed data and analysis from spreading?

AI in science and publishing is in its early stages, but it is already a reality that needs to be addressed and further developed.  Together, we should provide an open path forward to the potential of AI in scientific innovation.

The content generated on this blog is for information purposes only. This Article gives the views and opinions of the authors and does not reflect the views and opinions of the Impact of Social Science blog (the blog), nor of the London School of Economics and Political Science. Please review our comments policy if you have any concerns on posting a comment below.

Image Credit: Google DeepMind on Unsplash.