I finished my PhD before generative AI (genAI) became common, and I wrote my thesis without using it. By the time I started using these tools, I had already spent years reading papers, designing experiments, and learning to question results. That did not make me immune to mistakes, but it gave me experience and scientific judgment to compare AI answers against.
How can you build scientific judgment while using AI? Whether you are starting your research career, supervising students, or already established but unsure how to use AI, this article offers practical ways to use these tools while continuing to develop your own critical thinking. The focus is on questioning AI suggestions, checking them against evidence, and discussing uncertain points with experienced colleagues.
Plan Your Experiments
Learning how to design experiments independently is an important research skill. Start with your research question, then think about which experiment could answer it and which controls you would need to interpret the results. This learning process does not stop when you finish your PhD. Even senior researchers discuss experimental designs with their colleagues and supervisors. Working through these decisions is part of developing your scientific judgment.
Design Your Experiments Before Asking AI
If you ask AI to design the experiment first, you may spend less time practicing those decisions yourself. Instead, if you want to use AI, try designing the experiment first and then ask AI to challenge your design or identify a missing control. If it suggests a change, do not simply copy it. Ask yourself: Does this make sense? What would this control tell me? Why did I miss it, and what should I consider next time? If you are a supervisor, you can also encourage your students to explain why they accepted or rejected a particular AI suggestion, to see their reasoning.
Choose a free resource to help you move forward
EBOOK
Guide to Special Stains for Histology
DOWNLOAD
SMART Experiment Template
And, of course, discuss the design with your colleagues or supervisor before performing it at the bench. This way, you stay involved in the decisions, with AI as an additional tool.
Check Your Calculations with Established Methods
Once you have a design, you also need to work out the practical details. Working in the lab brings plenty of calculations. Even if math was never your favorite subject, it is important to understand how to calculate cell seeding numbers, dilution factors, and reagent concentrations. I would recommend working through the calculation yourself first. You can then check it with an established calculator suited to the task: for example, a mass–molarity calculator[1] for preparing a solution from a weighed compound, or a dilution calculator for diluting a stock solution.[2] Check the units, stock and final concentrations, and final volume.
Then, you can check calculations with AI, and it can help you find a possible mistake. However, AI’s answer should not be your only check. The same principle applies to analyzing sequences, designing primers, and so on. Use established tools, and use AI only to help you understand how those tools work or to explore possible approaches; check the settings and outputs yourself.
Check Protocols Against Reliable Sources
The protocols you will use need the same attention. If AI helps you find or summarize a protein pull-down or DNA/RNA extraction protocol, go back to the source before using it. This might be the paper and its supplementary methods, a validated protocol from your lab, or the relevant manufacturer’s instructions. Check the reagent concentrations, incubation times, and temperatures. Should the step be done on ice, at 4°C, or at room temperature?
In my own experience, AI has changed protocol steps, including temperature conditions. If you regularly prepare lysates for Western blotting, keeping samples cold during specified steps may feel obvious. But a beginner may not know why those conditions matter, and AI may miss or change them; for this reason, protocols should be used based on correct sources.
At first, checking the papers you intend to rely on may feel slower than asking AI for a summary and immediately planning an experiment. But an incorrect protocol can cost you time and reagents if you have to repeat the experiment. Carefully reading the relevant methods is part of preparing the experiment.
Follow Scientific Claims Back to the Evidence
What about brainstorming? When you are starting a research project, discussing ideas with AI may feel easier than bringing every question to your supervisor or colleagues. You can explore an idea without worrying that your question is too basic. That is fine, but you still need to check what comes back.
AI and Hallucinations
GenAI can produce plausible-sounding but incorrect information, often called hallucinations.[3] This can include invented references or citations to real papers that do not support the claim being made.
This happened to me when I was studying two proteins and trying to understand whether they had different molecular interactors. While brainstorming with AI, it suggested an interactor I had never heard about. It gave me a couple of papers that supposedly supported this interaction for my protein of interest, but when I checked those papers, some did not exist, and others were about something completely different.
However, finding a real paper is only the first check. AI may misinterpret its findings or describe the conclusions more confidently than the evidence allows. When you are unfamiliar with a topic, you might recognize that the source is trustworthy but still miss that the AI’s explanation does not match the results.
Five Questions to Ask Before Trusting an AI Answer
So, how can you check an answer when you are still learning the topic? Try working through these questions:
1. Does the Paper Exist?
Find it independently, for example through PubMed or the journal website, and check the title, authors, and DOI.
2. Does It Support This Particular Claim?
Locate the relevant experiment or figure and read the results and methods behind it.
3. What Exactly Was Demonstrated?
For a protein interaction, was direct binding shown, or were the proteins found in the same complex? Check which cells, species, and experimental conditions were studied.
4. What Are the Limitations?
Compare the AI’s certainty with the limitations in the paper. Also check whether the authors’ conclusions are supported by their results or whether they have exaggerated their conclusions beyond what the results really support. AI may repeat the authors’ conclusions without pointing out that the results do not fully support them.
5. Who Can Help Assess It?
Bring the specific claim or paper to a colleague, supervisor, or journal club, and try to discuss whether the claim is supported by the data.
You do not need to know everything before you can use AI. But you do need a way to check the answers and recognize when you need help. Reading papers, comparing findings, and discussing them with other researchers are ways to practice this. Asking AI to be critical or provide sources can be useful, but it does not replace these checks.
Effective Usage of AI
Ask Why
So, where can AI be useful while you are learning? One example is asking “why” a particular reagent or step is included in a protocol.
For instance, you could ask why sodium orthovanadate is added to a Western blot lysis buffer (It inhibits protein tyrosine phosphatases, helping preserve phosphorylation during sample preparation). Asking for an explanation can help you understand the protocol, but check the answer against a reliable reference before relying on it.
Troubleshoot
Troubleshooting is another use I have found helpful. If an experiment doesn’t work, AI can offer possible explanations to investigate. In my own experience, some of its suggestions have been useful to try. The next step is to consider whether each suggestion fits your experiment and discuss important changes with someone experienced in the method.
Test Your Reasoning
You can also use AI to practice thinking. Ask for a hint, an alternative explanation, or a question that challenges your reasoning. Try answering first, then compare your answer with AI’s and work out why you agree or disagree. Afterward, try explaining the idea in your own words without looking at the answer.
Analyze Code
AI can also help with coding. For example, if you have images to analyze in Fiji/ImageJ, you can ask it to write or modify a macro for the analysis. A macro is a small script that automates a series of tasks, such as applying the same measurements to several images. Describe the task using a generic example, rather than uploading your research images. But you still need to understand what the macro does. A macro that runs without an error is not necessarily measuring what you intended. Test it on a small set of images and compare the output with results you have checked manually. If you are new to macros, ask an experienced colleague to help you review the code and its output.
Protect Your Research Data
Before putting unpublished data into an AI tool, check what your institute allows and whether the tool is approved for that type of information.
Some AI services use data that you’re putting in it to train or improve their models. However, every AI uses data differently, and it is too simple to assume that every AI tool uses your data in the same way. Similarly, paying for a license does not, by itself, tell you whether a tool is appropriate for your research data.
Training is only one part of the concern. You also need to consider how the service stores the information and who may have access to it. Research involving human participants, industrial collaborators, or potentially patentable findings may have additional restrictions on sharing data. Institutions also have their own guidance. For example, King’s College London requires doctoral students to discuss confidentiality, data protection, copyright, and ethical considerations with their supervisory team and receive approval before uploading research material.[4] Your own institute’s requirements may differ.
This does not mean that all AI-assisted data analysis is inappropriate. There is a difference between asking for help with analysis code and uploading the data itself. In the Fiji macro example above, you can keep your images on your computer and ask for code using a generic description. Just make sure the prompt and any code you share do not contain confidential information.
A Brief Example of my Current Workflow
In my daily routine, I use Perplexity to help find literature, Claude for help with data analysis, modifying Fiji/ImageJ macros and brainstorming, ChatGPT for literature search, brainstorming and language polishing, NotebookLM to quickly find specific information in my library of papers (that I have already read!), and Gemini for language polishing.
These are examples of what I currently use, rather than recommendations that one tool is better than another. I am still learning how to ask clearer questions and get more useful answers, and my workflow may change as the tools develop.
For literature searches, I still check the papers themselves. For analysis code, I check what it does and whether the output makes sense. The same checks apply no matter which tool I use.
Transparency matters too. Keep a record of which tools you used and what they helped you do, and check your institute’s and target journal’s requirements for acknowledging AI use.
Conclusion: What to Check Before Using an Answer
My main take-home message is that AI can help with many research tasks, but you should use it in a way that strengthens your understanding and scientific judgment.
Before using an answer, ask yourself: Do I understand the reasoning? Have I checked the relevant evidence or calculation? Am I allowed to share the information in my prompt? Do I need help from someone experienced?
You do not have to spot every mistake on your own. But keep asking questions, checking evidence, and discussing your decisions. Use AI to support that process, without letting it take over your thinking.
AI Disclaimer: ChatGPT was used to assist with language editing. All views expressed are the author’s, and the author takes full responsibility for the final content.
References
- Molarity Calculator. Merck KGaA. [Accessed 31 October 2026]
- Solution Dilution Calculator. Merck KGaA. [Accessed 31 October 2026]
- Miryam Naddaf et al. (2026) Hallucinated citations are polluting the scientific literature. What can be done? Nature 652(8108):26–29.
- King’s College London. Generative AI: Guidance for doctoral students, supervisors and examiners. Published October 2023. [Accessed 31 October 2026]
You made it to the end—nice work! If you’re the kind of scientist who likes figuring things out without wasting half a day on trial and error, you’ll love our newsletter. Get 3 quick reads a week, packed with hard-won lab wisdom. Join FREE here.

