I’ve had many researchers come to me during AI research consultations asking which tool they should use, but almost no one asks how to use it properly. That’s the question I want this column to answer. The tool matters less than what you do with it.

Biology already has its own well-known AI tools. Structure prediction models like AlphaFold have reshaped entire subfields. But this piece is about something more basic: a way to think through whatever AI tool ends up on your screen, so that using it stays a deliberate, responsible choice rather than becoming a habit you can’t quite explain.

A Five-Step Way to Think About It

Here’s a rough framework I keep coming back to: framing, tool selection, execution, verification, documentation.

Framing comes first. Decide what role AI should play, and what it shouldn’t, for the task in front of you. In research involving patient data or other sensitive information, that’s usually the deciding constraint, before any tool gets chosen. Tool selection comes next, and it must match the role you just defined. Then, execution is the actual drafting, prompting, back-and-forth. The last two steps are often overlooked, yet they’re the ones that matter most. Verification means checking what you got against primary sources and your own judgment. Documentation means writing down where AI was involved, clearly enough that someone else could follow what you did.

Let’s see how this plays out. First with reading and framing a question, then with writing.

Reading Papers, Framing a Question

Imagine you’re starting a new project with a rough direction or interest, not yet a real research question. AI can help you see themes across a stack of papers fast. But it can also make certain gaps look more important than they actually are, simply because they show up often in whatever the tools happened to pull. That’s the framing risk. The AI isn’t deciding your question, but it’s nudging which options feel worth considering.

For tool selection, Asta and Consensus are good starting points for finding papers. Both search semantically, meaning they find papers that use different words for the same idea, something keyword search misses constantly. NotebookLM works differently: once you’ve gathered a set of PDFs, it lets you ask questions across all of them, with every answer linked back to the source. Many of these tools also let you narrow what gets pulled in the first place, limiting results to peer-reviewed sources or a specific field, which cuts down on noise before you even start reading. None of these replace reading the papers. They just cut down the time it takes to find out which papers are worth reading.

Execution rarely means typing one perfect prompt and getting a finished question back. It’s a few rounds of back-and-forth. A broad interest narrows to a specific population, then a specific outcome measure, then something you could actually design a research question around.

Verification matters most here, and it’s not just about checking whether a source is real. AI summaries tend to drop exactly the parts that keep research honest: limitations sections, boundary conditions, the reasoning behind why a study picked one method over another. Read the actual papers behind any summary before you build a hypothesis on it.

Documentation can be one line in your notes, something like ‘AI helped surface certain papers’, or ‘AI helped narrow a broad topic into something specific’. If it shaped the direction of the project, it’s worth having some record of that, even as simple as a line in your notes.

The question you end up with is still yours. AI can help you brainstorm and sharpen it, but it doesn’t get to decide what’s worth asking.

Writing and Grant Proposals

Now let’s look at the same five steps in a writing scenario, though the risks shift. Framing means deciding upfront which parts you’re comfortable letting AI draft or polish, and which parts need your own judgment: specific claims, how you interpret your data, anything where getting the nuance wrong actually matters.

General-purpose tools like ChatGPT or Claude handle drafting and restructuring well. Grammarly still earns a place even if a chatbot wrote your first draft, since it catches sentence-level issues a generation tool won’t flag on its own. Academic-focused tools like Paperpal or Writefull are built specifically for manuscript and grant language, which helps if English isn’t your first language or a journal’s style conventions matter for a specific submission.

Execution is draft, revise, draft again, same as always.

Verification means checking every AI-assisted sentence for accuracy, not just tone. What ends up in your grant application is still your responsibility, in every sense that matters.

Documentation matters most in this framework, because funders and publishers are actively writing rules on AI use. NIH made that clear in a 2025 notice, NOT-OD-25-132: AI shouldn’t be generating your core scientific ideas, your hypotheses, or the key sections of a proposal. The general read across university guidance is that basic copyediting and language cleanup are fine. The scientific logic has to be yours. Journals have their own versions of this too, usually requiring disclosure for substantial AI use like drafting or analysis, while treating a quick grammar pass differently. Given how often these policies shift, it’s worth checking current guidance before you submit rather than relying on what was true last year.

The Tool Was Never the Point

The choice of tool matters less than it seems in the moment. The same five steps apply whether it’s a general chatbot or something built specifically for biology.

What actually separates responsible use from the rest comes down to one thing. AI makes every step faster. Faster isn’t the same as done right.

We put together a handbook at MIDAS that walks through where AI fits at each stage of a research project, from planning and writing like the examples here, to the more technical side of data analysis. If you want to go deeper, the MIDAS AI in Research Handbook is available online and free.

A note on AI use in writing this piece: I used AI tools to draft and polish the language. All ideas, framing, and technical claims are my own, and every AI-assisted sentence was checked against my own knowledge and the sources cited here before inclusion. This is consistent with the disclosure practices described in the article itself.

Reference

• Allen Institute for AI. Asta. https://asta.allen.ai/
• Consensus. https://consensus.app/
• Google. NotebookLM. https://notebooklm.google/
• OpenAI. ChatGPT. https://chatgpt.com/
• Anthropic. Claude. https://claude.ai/
• Grammarly. https://www.grammarly.com/
• Paperpal. https://paperpal.com/
• Writefull. https://www.writefull.com/
• DeepMind. AlphaFold. https://alphafold.ebi.ac.uk/
• MIDAS AI in Research Handbook. https://midas-ai-in-research.readthedocs.io/
• National Institutes of Health (2025). Notice NOT-OD-25-132. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html