Large Language Model (LLM)

A large language model (LLM) is a type of artificial intelligence system built on deep neural network architectures, typically transformers, trained on vast corpora of text data to generate, summarize, translate, and analyze natural language. LLMs such as the GPT, Claude and Gemini model families have rapidly entered the research landscape, where they are used for literature screening, data extraction, drafting, and analytical support.

LLMs are changing many research workflows quickly. They can process thousands of abstracts in minutes, generate first drafts of manuscripts, summarize complex papers, and extract structured data from unstructured text. These capabilities have the potential to accelerate systematic reviews, reduce the burden of repetitive research tasks, and democratize access to analytical tools. At the same time, LLMs raise serious questions about transparency, reproducibility, bias, accuracy, and authorship. Journals, publishers, professional organizations and funding agencies have issued policies on how LLMs may be used in research. The ICMJE recommendations and the COPE position statement on authorship and AI tools both say that AI tools cannot be listed as authors, that authors must disclose how they used them, and that the human authors remain responsible for the content; individual journals and funders add their own rules, which continue to change.

LLMs are built using transformer architectures that process text through layers of attention mechanisms, learning statistical patterns in language from training data that can include books, academic papers, websites, and other text sources. During training, the model learns to predict the next word or token in a sequence, and through this process it acquires a broad but imperfect representation of language, facts, reasoning patterns, and stylistic conventions.

When a researcher uses an LLM, they provide a prompt (a question, instruction, or passage of text), and the model generates a response based on its learned patterns. The response is usually generated by sampling from the model's probabilities rather than by a fixed rule, so the same prompt may produce different outputs on different occasions, and the model can generate text that sounds authoritative but is factually incorrect (a phenomenon called hallucination). LLMs do not "know" things in the way a database does; they generate plausible text based on patterns, which means they can confidently produce fabricated citations, invented statistics, or inaccurate summaries.

In research applications, LLMs are being integrated into several stages of the workflow. For literature screening, they can classify abstracts as relevant or irrelevant to a review question with accuracy that in some evaluations approaches that of trained human reviewers, though performance varies across reviews, prompts and models, so they still require human verification. For data extraction, they can pull structured information from study reports, though extraction errors are common enough to require manual checking. For writing, they can generate first drafts that researchers then revise, though concerns about bias, plagiarism, and the dilution of original thinking apply. For analysis, they can suggest statistical approaches, write code, and interpret results, though the researcher remains responsible for verifying the correctness of all outputs.

LLMs are most valuable as assistants for repetitive, time intensive tasks where human oversight remains in the loop. They are useful for initial screening of large numbers of abstracts, for generating draft text that will be substantially revised, for coding and data transformation tasks, and for exploring unfamiliar methodologies. They are not appropriate as the sole judge of inclusion decisions, as unsupervised generators of final manuscript text, or as sources of factual claims without independent verification. Researchers should disclose the use of LLMs as the target journal's policy and relevant professional standards require: COPE asks for disclosure in the methods (or a similar section), and the ICMJE asks for it in the cover letter and in the appropriate section of the submitted work.

The most dangerous mistake is treating LLM output as a reliable source of factual information without verification. LLMs can fabricate references, invent data, and produce plausible but incorrect statistical interpretations. Another error is failing to disclose LLM use, which raises concerns about transparency and academic integrity. Some researchers use LLMs to screen studies for systematic reviews without validating the model's accuracy against human reviewers on a pilot sample, which can introduce systematic bias into the evidence base. Finally, relying on LLMs to generate entire manuscript sections without substantial researcher input and revision can produce generic, superficial text that lacks the depth and nuance expected in scholarly writing.

A research team conducting a scoping review of 15,000 abstracts on digital health interventions uses an LLM to classify each abstract as potentially relevant or not relevant based on predefined criteria. In this illustrative example, the model achieves 94% sensitivity and 87% specificity when validated against a random sample of 500 abstracts screened by two human reviewers. The team uses the LLM's classifications as a first pass, with human reviewers examining all abstracts flagged as relevant and a 10% random sample of those flagged as irrelevant to estimate how many relevant abstracts the model missed (false negatives). The use of the LLM is disclosed in the methods section of the published review.