How Scientists Use AI: Tools Across the Research Cycle

Scientific research is a cycle: reviewing what is known, forming hypotheses, designing experiments, collecting and analyzing data, and communicating results. AI now touches nearly every stage of this cycle, which is why it has become one of the most consequential tools in modern research.

The reasons are the same forces at work in other professions, amplified by science's particular demands. Much of research is language-heavy (reading thousands of papers, writing manuscripts) and computation-heavy (analyzing data, writing code, running models), and both are areas where AI excels. A scientist can spend a huge fraction of their time on tasks that are essential but not the creative core of discovery, and AI can compress many of them.

But science has a distinctive requirement that shapes how AI must be used: rigor. Science advances by claims that are verified, reproducible, and honestly reported. A tool that produces plausible-but-unverified output sits in tension with a field whose entire value rests on truth and reproducibility. This makes AI in science genuinely powerful and genuinely perilous, and the two cannot be separated.

This cursus is a practical guide for researchers: the categories of AI tools across the research cycle (this lesson), the core workflows of literature review, data analysis, and coding (lesson two), and the integrity, reproducibility, and limits that responsible scientific use demands (lesson three). The aim is to help scientists accelerate their work without compromising the rigor that makes it science.

AI tools for scientists span the research cycle. It helps to see them by the stage they serve.

• Literature tools: search, summarize, and synthesize the scientific literature, finding relevant papers, extracting key findings, and mapping what is known on a topic far faster than manual reading.
• Data analysis and coding tools: write and debug analysis code, perform statistics, wrangle and visualize data, and explain results, acting as a tireless programming and analysis assistant.
• Writing tools: draft and refine manuscripts, abstracts, grant proposals, and responses to reviewers, and improve clarity, especially valuable for non-native English writers.
• Hypothesis and idea tools: brainstorm hypotheses, suggest experiments, and surface connections across fields, as a thinking partner rather than an authority.
• Specialized scientific models: purpose-built AI systems that make actual scientific predictions in a domain, such as predicting protein structures, properties of materials, or molecular interactions.

The first four categories are general assistants applied to research tasks; the fifth is fundamentally different and deserves its own step. The unifying theme for the general tools is familiar: AI accelerates the supporting work of research, reading, coding, writing, brainstorming, so the scientist can focus on the design, interpretation, and judgment that constitute actual discovery.

And the familiar framing applies with extra force in science: these tools produce a draft or a suggestion, not a verified result. A summarized finding must be checked against the paper; generated code must be validated; a suggested hypothesis is only a starting point for real investigation. In a field defined by verification, the scientist's checking is not optional polish but the essence of the work.

A distinction matters more in science than almost anywhere: general-purpose AI assistants versus specialized scientific models. They are different kinds of tools with very different reliability.

General-purpose assistants are the broad language models, applied to research tasks like summarizing, coding, and writing. They are versatile and useful, but they carry the standard risks: they can hallucinate facts and citations, they are not authoritative sources of scientific truth, and their output always requires verification. They help you do research faster; they do not know science.

Specialized scientific models are a fundamentally different category: AI systems trained for a specific scientific prediction task, often producing results that are genuinely novel and validated. The landmark example is AlphaFold, developed by DeepMind, which predicts the three-dimensional structure of proteins from their amino-acid sequence, a decades-old grand challenge in biology. It has predicted structures for hundreds of millions of proteins, transforming structural biology, and the work was recognized with a share of the 2024 Nobel Prize in Chemistry. Similar specialized models exist for materials, molecules, and other domains.

The key insight: these are not the same thing, and conflating them is a mistake. A general chatbot's confident biology claim is unverified text; a specialized model's protein-structure prediction is a scientific tool validated on real data (though still confirmed experimentally where it matters). Both are useful, but knowing which kind of tool you are holding, and its distinct reliability, is essential to using it responsibly.

The most dangerous failure mode for scientists using general AI is the same one that endangers lawyers: hallucinated references. General language models can invent citations, papers, authors, journals, findings, that do not exist but look completely real.

In science this is especially insidious because citations are the currency of credibility. A scientist who asks an AI to "find papers supporting this claim" or to "write a literature review with references" may receive a list of plausible-looking citations, some of which are entirely fabricated: real-sounding titles attached to real-sounding authors in real journals, describing studies that were never conducted.

The consequences in science are serious. A fabricated citation in a manuscript undermines the paper's integrity, can mislead other researchers, and, if it reaches publication, damages the scientific record and the author's reputation. Unlike a general error, a fake reference actively corrupts the network of verified knowledge that science depends on.

Why it happens is the now-familiar mechanism: a general model generates statistically plausible text, and a citation that fits the pattern of a real one is exactly what it produces, whether or not the paper exists. The model does not consult a database of real papers; it predicts what a citation would look like.

The rule for scientists is therefore absolute: every citation and factual claim from a general AI must be verified against the actual source before it is used. Never cite a paper the AI mentioned without finding and reading the real paper. Specialized literature tools that retrieve from real databases reduce this risk, but the scientific duty to verify sources, always central to research, applies with full force to AI output. In science, an unverified citation is not a shortcut but a landmine.

Science has a defining standard that shapes how AI must be used: reproducibility. A scientific result must be reproducible, others must be able to repeat the method and get the same result, and AI can quietly threaten this if used carelessly.

The tension is that AI outputs can be variable and opaque. Ask a general model the same analysis question twice and you may get different answers; its reasoning is not transparent; and if AI is woven into an analysis without careful documentation, another researcher may not be able to reproduce exactly what was done. A result that cannot be reproduced is not yet solid science.

This imposes specific discipline on scientific AI use:

• Document AI use in methods. If AI contributed to an analysis, that should be recorded transparently enough that the work can be understood and repeated, which is increasingly expected by journals.
• Prefer deterministic, inspectable steps for analysis. Where AI writes analysis code, the code itself, not a hidden AI judgment, is what runs, and it can be shared and rerun. The reproducible artifact is the code and data, not the chatbot conversation.
• Do not let AI be an unexaminable step. An analysis whose crucial step is "the AI decided" is hard to reproduce or defend; the reasoning must be made explicit and checkable.

The deeper principle is that reproducibility requires transparency, and AI's opacity is in tension with that. The responsible approach uses AI to help produce transparent, reproducible work (writing shareable code, drafting documented methods) rather than as a hidden oracle whose outputs cannot be traced. Used this way, AI accelerates research while preserving the reproducibility that makes results trustworthy, a theme the final lesson develops into full scientific-integrity practice.

Step back to a balanced view of what AI actually changes for research, avoiding both the hype of "AI will make discoveries for us" and the dismissal of "it is just a chatbot."

What AI genuinely changes: the speed and reach of the supporting work of science. Literature that took weeks to survey can be mapped in days; analysis code that took hours to write can be drafted in minutes; manuscripts can be clarified and polished faster; and specialized models can make predictions (like protein structures) that were previously impossible or slow. This is a real acceleration of the research process, and in the case of specialized models, a genuine expansion of what is scientifically achievable.

What AI does not change: the scientific method itself. The design of a rigorous experiment, the interpretation of results, the judgment about what a finding means, the honesty of reporting, and the responsibility for the work all remain human. AI can suggest a hypothesis, but testing it properly is science; AI can run an analysis, but deciding whether the result is real and what it implies is the scientist's job. Critically, AI does not verify truth, science does, through experiment, replication, and peer review.

So the realistic picture is that AI is a powerful accelerator and, in specialized forms, a genuine instrument of research, while the core of science, rigorous method, honest interpretation, and reproducible verification, remains firmly human. Scientists who benefit most use AI to move faster through the supporting work and to access new predictive tools, while applying undiminished rigor to everything AI produces.

That balance, real acceleration paired with unrelenting verification, is the theme of this cursus, and it sets up the practical workflows and the integrity practices of the next two lessons.

General AI accelerates the supporting work at each stage of research, literature, analysis, coding, writing, while specialized models make domain predictions; all output passes through the scientist's verification, and the scientific method stays human.