In 2024, almost two years ago, the first systematic, large-scale analysis on the prevalence of large language models (LLMs) in published research was published.

The results were impressive. Even by this point in time—just about one year after the public release of OpenAI’s GPT 4.0 and Anthropic’s Claude—scientists made surprisingly frequent use of the new technology in their research papers. The numbers ranged from 6.3% of mathematics papers up to 17.5% of computer science papers. Given the average turnaround time of research studies (about ten months in STEM fields), it’s an impressive result that some 10% of research studies were actively making use of the groundbreaking and newly available artificial intelligence technology.

Source: Nature.

The proliferation of AI in science has continued unabated in the two years since. A 2025 survey published by Frontiers found that more than half of academics used AI tools while peer reviewing manuscripts—often against journal guidance or regulations. This was a 24% increase over the prior year.

70% of surveyed academics use AI for writing papers, and nearly a quarter use AI for analysis, design, or methodology. Despite this, over 60% of researchers said that they have concerns that AI tools are being misused by researchers—and over 50% have actually observed AI misuse.

Scientists believe using AI in research has mixed results. On the plus side, 63% of peer reviewers believe that AI enhanced the overall quality of the research they were reviewing. Yet a majority also said that AI tools introduced errors into the manuscript and made them doubt the integrity of the research.

The position of AI in research in 2026 is a curious one. While mistrust and skepticism remain, researchers are increasingly relying on the technology—in particular to draft reports, summarize findings, and flag research misconduct. The Frontiers survey concludes that by and large, researchers have not made use of truly transformational applications.

In this article, we’ll cover the common as well as the most outstanding use cases of AI in science as of early 2026. The impact of generative AI has attracted the eyes of the scientific world, and deserves the careful scrutiny that we will exercise in this review.

General applications

As the Frontier survey revealed, most scientists are, in fact, using AI. But rather than applying LLMs to the core meat of their research like optimizing methodology, performing data analysis, or running experiments, they’re using AI here and there, on the fringes. Especially, they’re using AI for manuscript review and for writing.

However, a lot of scientific journal guidance warns strictly against using AI for writing. Using AI to write core content of researcher papers has been widely shown to generate false citations, introduce inconsistencies, and more—not to mention, undisclosed AI writing can and sometimes will be considered plagiarism. Here’s a list of common dos and don’ts for AI in paper writing:

• Don’t: Copy and paste AI-generated text.
• Don’t: Use AI citations without checking the sources.
• Do: Use AI to find and improve on phrasing.
• Do: Use AI to get feedback.

Probably the second biggest application of generative AI in scientific research is in coding. LLMs function as coding copilots, in a similar way that they can function as writing copilots. Experts strongly advise againstcopying and pasting code from LLMs. On the other hand, in fields like data science and astronomy—which require extensive coding that may be outside of a scientist’s direct area of expertise—scientists can greatly benefit from help with AI-generated code.

For example, in data science, using generative AI can help analysts rapidly explore various models quickly and effectively. A recent peer-reviewed study shows a wide variety of current applications by scientists, such as writing code in a new language for the first time. One of the most frequently observed uses in the study was when scientists were seeking information about the methods available within programming libraries, example use cases for those functions, and the meanings of unfamiliar functions.

Perhaps unsurprisingly, a dos and don’ts list for AI in coding looks quite similar to the writing list:

• Don’t: Use AI to write the entire code for a project.
• Don’t: Use AI-generated code without understanding and testing it.
• Do: Use AI to narrow down sources of bugs.
• Do: Use to generate chunks of code that can be tested and understood.

Powerful applications

At present, many (and perhaps even most) scientists have integrated AI to some degree in their workflows, primarily to assist with writing, collecting information, and coding. More groundbreaking applications have proved to be fewer and far between. So here are four examples, across disciplines, of scientists taking full advantage of LLMs to achieve novel scientific discoveries.

1. Predictive modeling

In the field of chemistry, LLMs are being employed to predict various chemical properties, from solubility to electron orbital energy difference (HOMO-LUMO gaps). This feat is made possible by using the language-interfaced fine-tuning framework (known as the LIFT framework). LIFT treats scientific data as language, i.e. molecules as strings of text, training LLMs to be able to predict molecular structures.

One recent applications of this approach is predicting atomization energies. A team of researchers at the University of California Berkeley and the University of Madison showed that the LIFT framework yields good predictions.

2. Experiment design and execution

CRISPR gene editing. Source: Iota Sciences

Studies show that automated experiment design and execution has several benefits compared to hard-coded planners, in particular being more flexible in unexpected situations.

One promising example in this domain is CRISPR-GPT. The program uses LLMs to automate the design and execution of gene-editing experiments, including selecting CRISPR systems, designing gRNAs (telling CRISPR system where to make a cut), drafting protocols, and analyzing data. The study shows that the program can help non-experts conduct gene experiments potentially relevant to their core field.

A similar example in the field of genetic perturbation experiments, which study gene fuction is the BioDiscoveryAgent. This software likewise presents an interface to execute perturbation experiments from start to finish.

3. LLM software interfaces

Chemical software—such as for visualizations—can be incredibly difficult to use, and require up to days of time investment to learn how to use even simple visualization tools. To address this issue, a team of scientists at the University of Pittsburgh published an example of how LLMs can write code for existing visualization tools to make them easy to use and apply.

Rather than learning the specifics of interacting with novel software, users can simply input queries to an LLM, which will then manipulate the software. Other published research has made use of similar applications, including creating an LLM-interface for protein engineering tools.

4. Reusing Datasets

Jeremy Magland’s lab at the Simons Center is using AI to reuse complex, large datasets from the Distributed Archives for Neurophysiology Data (DANDI) Archive. DANDI hosts hundreds of datasets of everything from recordings of brain activities to behavioral and stimulus data. The AI research agent explores data sets to produce automated reviews that can alert scientists if a dataset has a potential discovery worth working on.

These examples make it clear that the best and most effective uses of AI are those that enable new scientific discoveries. By analyzing massive datasets, machine learning models allow for previously impossible scientific discoveries. These uses are clearly more transformational than simple efficiency-boosts at the level of writing and coding.

Literature discovery

One of the biggest applications of artificial intelligence—but also one relatively lesser explored by scientists compared to writing and coding—is literature discovery.

Since 2022, over 5 million scientific articles have been published every year. That results in massive number of yearly articles even in niche disciplines—somewhere between 130,000 and 250,000 per discipline (like neuroscience), and between 300 and 1,500 papers in a typical niche scientific specialty (such as synaptic plasticity or visual cortex processing within neuroscience). That’s simply way too many articles and discoveries for any individual to keep up with!

Number of scientific papers. Source: Reddit (Bornmann et al)

Artificial intelligence, especially gen AI, can help solve this volume problem in a few ways. LLMs—especially third-party plugins for GPT like ChatPDF.com, UseChatGPT.AI, GPT Academic, SciSpace copilot, and more—can generate summaries from open-access scientific articles. Certain structured prompts can be used to tailor AI responses to specificy format and data-points for the summaries, meaning that scientists can have AI essentially pre-read articles for them.

LLMs can also be used to select articles to read. Natural language processing algorithms can scan large volumes of research papers, and from there extract key findings, identify trends, and summarize insights. NewsRx is developing one such tool that generates concise summaries of scientific article abstracts. This brand new tool delivers abstracts to scientists and allows them to quickly determine whether or not an article is worth a read.

Lastly, while it may feel that traditional data sources already provide an overabundance of information, LLMs solve another key problem in the scientific ecosystem: inclusive data. Traditional databases tend to neglect “grey literature,” or resources and reports issued by governmental and nongovernmental organizations rather than scientific organizations. Many fields, such as conservation and ecological restoration, can greatly benefit from making use of these data sources, which are excluded from typical scientific databases.

Conclusion

About two years ago, we summarized the contemporary scientific approach to LLMs. At the time, ChatGPT was emerging as a tool to automate simple tasks, like summarizing topics and composing emails. Meanwhile, major scientific journals were moving to ban AI-written studies, and common versions of LLMs didn’t even have full access to the Internet yet.

In the two-plus years since then, LLMs have grown markedly more powerful, and their usage has disseminated widely among scientists. And yet the vast majority of researchers still only use them for routine tasks. But as powerful, disruptive applications bubble up one study at a time, the true disruptive, transformative potential of the technology will begin to come into full view.

A Note of Caution

No matter your use of AI, make sure to follow these simple ethical safeguards:

1. Follow the rules of your target journal and avoid plagiarism. Every journal has a slightly different guideline for AI tools, and treats AI-generated text in slightly different ways.

2. Outline the relevant risks before you use an LLM. Since AI can both benefit and harm studies, scientists should only proceed with making use of it if potential risks can be mitigated. This is easy in cases of factual errors that can be corrected through proofreading, but correction can be impossible in instances such as biases in a model’s underlying training data.

3. Respect confidentiality. Conversations that happen in many LLM models, including ChatGPT, are used as training data. This means that typing in medical records, for instance, would result in a breach of confidentiality.

4. Verify the truthfulness of AI-generated content and AI-located sources and references. This is a be-all and end-all for LLM-powered research!