Careful use of a reference manager can help scientists to keep control of their to-read pile. Credit: Impact Photography/Shutterstock
In the past 13 years, I have added 10,000 papers to the reference manager that I first used when I was a PhD student. The topics span Alzheimer’s disease, genetics, biostatistics, epidemiology, the philosophy of science and more. Some of these works taught me about my own field; others pulled me away from it, informing my understanding of other areas of science, alternative methods and fresh ways of seeing a problem.
That range reflects how I learnt to read the literature as a scientist. Early on in my PhD, during which I studied the genetic and environmental factors that contribute to cognitive decline, I thought I needed to read every paper closely and critically, especially if I planned to cite it. That belief turned my reference manager into an ever-expanding to-read list. I quickly realized, while writing my thesis proposal, that I could not read every paper closely. Part of my development as a scientist has been learning to read selectively, to the depth required for the task at hand.
Even selective reading, however, cannot fully relieve the pressure of keeping up with the exponentially growing scientific literature. Artificial-intelligence tools built on large language models (LLMs) promise fresh ways to search, summarize and sort the literature. When I first used ChatGPT, I was impressed by how well it could summarize scientific topics. But I soon learnt that it could also fabricate citations, and hallucinated references have become a visible sign of uncritical AI use in manuscripts.
Although the most recent LLMs, with built-in search, return real papers and summarize the literature more reliably, they do not eliminate the risk of uncritical citations. This is a modern version of an old problem, in which scientists cite papers they have not read, or have misunderstood or inherited from other reference lists. A key part of developing as a scientist today is learning how to use AI tools for literature management without letting them replace scientific judgement. This is also how I frame these tools for my students and trainees.
A workflow for reading the literature
In my own work, and when conducting training, I treat reading as a set of decisions about purpose, relevance and depth. AI can assist with those decisions, but it cannot eliminate them. Those decisions usually fall into four broad categories.
Discovery. Many papers that I read reach me before I go looking for them. This is passive discovery. Each morning, I check an RSS feed that pulls works from 35 journals, including specialist publications such as Alzheimer’s & Dementia and multidisciplinary ones such as Nature and Science. This gives me a regular view of what is being published across my main areas of interest. Newsletters, citation alerts and recommendations from colleagues serve a similar function, and social media can also reveal which papers are gaining attention across the scientific community.
Active discovery, by contrast, uses a search strategy to address a specific research need, such as identifying papers that inform a particular idea I want to develop. In the past, this meant using keywords or search terms in bibliographic databases such as PubMed or Google Scholar, and following citations from relevant articles. AI tools now add another route. When I do not know the right search terms, I increasingly use chatbots to do a first pass, asking them to identify potentially relevant papers across databases and the web. For example, for a grant application about biological ageing in Alzheimer’s disease, ChatGPT identified 52 potentially relevant papers across several searches; I added around half of these to my reference manager, and ultimately cited nine in the application.
Triage. The discovery stage flags many more papers than I can read. My RSS feed alone surfaces dozens of papers a day, and a broad PubMed search can return hundreds more. Reviewing titles and abstracts allows me to determine which papers to ignore.
Interesting papers that are not immediately useful go into my reference manager, tagged by topic so I can find them later. For papers that are relevant to an ongoing project, I copy their citation details into my project notes, grant manuscript or presentations. AI tools can increasingly help with triage, especially when a web or PubMed search returns more papers than I can reasonably inspect. I use LLMs to review titles and abstracts from exported search results, comparing them with the questions I am trying to answer and flagging papers that seem worthy of a closer look.
Skimming. Once a paper passes triage, I skim it to build a quick, high-level understanding. The goal is to identify the research question, main assumptions and key findings. I use the paper’s structure to guide my reading, focusing on subsection headings, topic sentences, the introduction and the conclusion. My notes at this stage are minimal, usually limited to copying useful passages directly into project notes or outlines for manuscripts and grants.
Skimming can be helpful for several tasks. When I am surveying papers from a new field, it helps me to build enough understanding of the background to ask focused questions. When I am looking for a citation for a particular piece of information, such as the diseases that have been associated with a gene, skimming helps me to locate the relevant passage and evaluate how robust the evidence is for that claim. I have also used AI tools at this stage, usually by uploading several papers and asking the tool to extract specific points relevant to my question. As with checking citations from published articles, I still use skimming to verify that AI-generated summaries reflect what the paper actually says.
Deep reading. For papers I need for research, grant writing or teaching, I read more actively and start making detailed notes. In essence, deep reading means understanding the paper’s research question, study design, main findings and interpretation well enough to explain it to someone else. It also means working through the authors’ argument closely enough to test their conclusions, identify their assumptions and connect the paper to the literature. This level of reading can take anywhere from 30 minutes to several hours, depending on how technical the paper is and how much I need to rely on it.
