How to Use AI for Data Analysis in Postgraduate Research

TL;DR:

• AI can help a Masters or PhD student clean data, write analysis code, explain methods and challenge a draft interpretation. You still choose the method, run the analysis, check the result and write what it means.
• The safest set-up keeps your data on your own computer. AI writes code using invented sample rows, and you run that code on the real file.
• Names, phone numbers, voice recordings and raw transcripts should stay out of public AI tools. Check your consent form, ethics approval and NACOSTI conditions before you upload any research data.
• Research on chatbots shows they handle textbook statistics well and are weaker at judging whether a method fits your data. That judgement is what examiners will test.
• Keep an AI analysis log from day one. It protects you when your supervisor or panel asks what the tool did.

Who this guide is for

This guide is for Masters and PhD students who have data, or will soon, and want to use AI without damaging their research. It covers survey data, interviews and secondary datasets. You do not need to be good at statistics or coding to follow it.

We have written about where the integrity line sits in postgraduate AI use in a separate post. This one goes into the practical work of analysis, step by step.

Your university's rules come first. If this guide and your graduate school disagree, follow your graduate school.

What can AI safely do at each stage of analysis?

AI is most useful on the mechanical work and as a critic of your thinking. The risk rises as you get closer to deciding what your data means.

The pattern is simple. Let AI speed up the typing and the checking. Keep the decisions.

How good is AI at statistics?

The honest answer is that AI does well on clear textbook problems and less well on judgement about real data. Three studies show the range.

• Textbook questions. A 2025 study in Cureus gave six free chatbots 20 hypothesis-testing scenarios. All six picked the right test every time. The test ran on a single day and used free versions only.
• Judging whether a method fits. The StatQA benchmark, a preprint from 2024, has 11,623 examples, and most of its tests used a subset of 1,163. The best model used without extra training, GPT-4o given domain knowledge, scored about 65%. The most common mistake was choosing a method whose conditions were not met. Models have improved since then, so treat the number as history and the type of error as the lesson.
• Coding interviews. A short 2024 workshop study compared GPT-4 with human coders on German-language data from an interview study with 47 people. Agreement was high on simple labels, such as naming a device, and only moderate on the most complex task. The authors advise checking AI coding task by task.

Put together, a chatbot may name the right test for a tidy example and still miss that your data breaks the test's conditions. Your own dataset is the hard case, so check its conditions yourself.

What are the rules on research data in Kenya?

Three layers apply: your research licence and ethics approval, the Data Protection Act, and your university's rules. Check all three before any data leaves your computer.

Research licence. NACOSTI says anyone intending to do research in Kenya must apply for a research licence. Read your licence conditions and your ethics approval. They may limit who can handle your data.

Consent. Read the consent form your participants signed. If you told them only the research team would see their answers, uploading those answers to an outside AI company may break that promise.

The Data Protection Act. Section 2 lists sensitive personal data. It includes health status, ethnic social origin, belief, marital status, family details, and sex or sexual orientation. Many social science and health studies collect exactly this. Three more sections matter:

• Section 49 says sensitive personal data can be processed outside Kenya only with the person's consent and confirmation of appropriate safeguards. Many AI tools are hosted abroad, so ask where your tool runs.
• Section 53 gives research some exemption, on conditions. One is that results are not published in a form that identifies people.
• Section 2 defines pseudonymisation. Data is only pseudonymised if the extra information that links it back to a person is kept separately and protected. Replacing names with codes while the key sits in the same folder does not meet that description.

Your university. We could not find a published AI rule for theses from the University of Nairobi or Kenyatta University when we checked on 9 November 2026. One may still exist inside a faculty. A student guide from August 2026 reports that the Commission for University Education has not issued a binding national rule on AI tools. Ask your graduate school in writing and keep the reply.

What the tool does with your data. Terms differ by tool and by plan. As one example, Google says that when Keep Activity is on, a subset of Gemini chats is read by human reviewers, and reviewed chats can be kept for up to three years. Read the current privacy page for the exact tool you use.

Use this table as a quick guide.

One warning about the amber zone. Removing names may not be enough. A head teacher in a small named sub-county, aged 58, with a rare subject combination, can be recognised without a name.

The safest way to work: bring the code to the data

Keep the real data on your own computer, and let AI work from a description of it. This one habit removes most privacy risk and also makes your analysis easier to repeat.

1. Freeze the raw file. Save a read-only copy and never edit it.
2. Write a data dictionary. List each column, what it means and which values are allowed.
3. Make a fake file. Type 10 invented rows with the same columns.
4. Ask AI for code using only the dictionary and the fake rows.
5. Run the code yourself on the real file, in R, Python, jamovi or the software your department uses.
6. Check by hand. Confirm the row count, one group's count and one mean against the raw file.
7. Log it. Record what you asked, what you used and how you checked it.

If your data is too sensitive even for that discussion, you can run an open-source AI model on your own computer, so the text stays on your machine.

A worked example, with invented data

Here is how this works from start to finish. Every number below is invented for teaching. A student surveys 240 smallholder farmers across two counties. The question is whether attending a training session goes with using a phone-based weather advisory.

Step 1: cleaning catches a silent error

The student asks AI for code that lists every unique value in each column. Two problems appear.

• County is spelled three ways: Kiambu, KIAMBU and Kiambu County.
• Twelve people have an age of 99, which was the code for "refused to say".

If nobody catches the second problem, mean age comes out as 43.9. Once those 12 rows are treated as missing, the mean of the other 228 people is 41.0. No error message appears. The number is just wrong.

Step 2: the first result looks exciting

The student runs a chi-square test on two yes-or-no columns.

The result is χ²(1, N = 240) = 9.70, p = .002. A chatbot asked to interpret this might write: "Training significantly increases adoption by 20 percentage points."

That sentence claims cause. A cross-sectional survey cannot show cause. It also ignores who chose to attend.

Step 3: the student asks a better question

The student knows the field and suspects smartphone ownership matters. So the student splits the table.

Among smartphone owners there is no difference at all. Among people without one, the gap is small and goes the other way, and it is not statistically significant (χ² = 0.65, p = .42).

Look at who attended. Of the 120 people trained, 90 owned smartphones. Of the 120 not trained, only 40 did. The 20-point gap mostly reflects who came to the training.

Step 4: the student writes the finding

Advisory use was higher among farmers who attended training (65%) than among those who did not (45%). However, training attendance was strongly related to smartphone ownership. Within smartphone-ownership groups, the difference was absent or small. These cross-sectional data do not show that training increased advisory use.

A full analysis would go further, perhaps with logistic regression. That is a conversation for you and your supervisor. The point is that a test result alone gave a tidy, confident and misleading sentence. Knowledge of the field found the problem.

Five prompts you can copy

These use the four Cs we teach: Clarity, Context, Constraints and Customisation. Replace the square brackets. Do not paste real participant data.

1. Stress-test your analysis plan

I am a [Masters/PhD] student in [field]. My research question is [question]. My design is [cross-sectional survey, n = ...]. My outcome is [variable and type]. My predictors are [variables and types]. I plan to use [method]. Do not suggest a new plan yet. List the assumptions of my method, how I can test each one, and three reasons a careful examiner might reject this choice. Use plain English and name a reference book or paper I can check for each point. If you are not sure a reference exists, say so.

2. Write cleaning code from a data dictionary

Here is my data dictionary and 10 invented rows: [paste]. Write [R/Python] code that reads my file, lists all unique values in each categorical column, flags values outside the allowed range, and counts missing values per column. Treat 99 in age as missing. Do not delete any rows. Save the cleaned data as a new file and leave the raw file unchanged. Add a one-line comment above each step so I can explain it to my supervisor.

3. Challenge your interpretation

Below are my results table and the paragraph I wrote interpreting them. All details are de-identified. Do not rewrite the paragraph. Tell me: (1) where I claim more than the design allows, (2) other explanations I have not considered, (3) any number in the paragraph that does not match the table, and (4) what variable I might need to check. Results and paragraph: [paste]

This order matters. You think first, and AI argues with you afterwards.

4. Practise your defence

Act as a strict external examiner in [field]. Here is my methods summary: [paste]. Ask me one question at a time about why I chose this method, how I checked its assumptions and what my limitations are. Wait for my answer. After each answer, tell me what was weak. Do not give me a model answer until I have tried twice.

5. Test a qualitative codebook

Here is my draft codebook and three invented excerpts in the style of my interviews. Apply the codes and show where two codes overlap or where a definition is unclear. Do not suggest final themes. I will develop the themes myself from the full transcripts.

How should you handle interviews and transcripts?

Treat AI as a junior helper on qualitative work, and keep reading your own transcripts. Interpretation is the method in most qualitative research, so a summary cannot stand in for it.

Transcription needs special care. A 2024 study of the Whisper speech-to-text system found that about 1% of its transcriptions contained whole phrases or sentences that were never spoken. A similar risk appeared in social work: a 2026 Ada Lovelace Institute report records a social worker who found that an AI summary wrongly indicated suicidal thoughts.

Many Kenyan interviews move between English, Kiswahili, Sheng and a local language. We could not find solid published accuracy figures for this kind of mixed speech, so test it yourself. Take a two-minute recording made with a colleague's consent, run it through the tool, and compare it word by word. For the real recordings, check your consent terms first and read every transcript while listening to the audio.

What should you disclose, and how?

Disclose what AI did in your analysis, and keep records that prove it. Follow your university's format where one exists.

The European research guidelines published in October 2026 are a useful model, although they do not bind Kenyan universities. They say researchers stay responsible for their output and that AI cannot be an author. They treat basic editing as minor. Using AI to interpret data analysis counts as substantial use that should be disclosed. They also ask researchers to describe how they checked AI's work.

For journal articles, the ICMJE recommendations say authors must disclose AI use at submission and remain responsible for the content. Failure to disclose may lead to corrective action.

Keep a simple log as you go.

A short disclosure could read: "I used [tool, date] to help write R code for data cleaning and to critique a draft of my interpretation. All analyses were run and checked by me. I chose the methods and wrote the interpretation. No participant data was entered into the tool."

Only write that last sentence if it is true.

What goes wrong most often?

Most problems come from trusting a result that looks finished. These are the ones to watch for.

On the p-value point, the American Statistical Association's 2016 statement is worth reading. It says a p-value does not measure the size of an effect or the importance of a result. It also calls for full reporting and transparency.

Can you answer these eight questions at your defence?

A panel tests whether you understand your own analysis. If you can answer these without notes, you are in good shape. For wider preparation, see our guide on how to prepare for your defence.

• Why did you choose this method over the obvious alternative?
• What assumptions does it make, and how did you check each one?
• How did you handle missing data, and what would change if you had handled it differently?
• Which cases did you exclude, and why?
• What does this coefficient or percentage mean in plain words?
• How large is the effect, and does it matter in practice?
• What else could explain this result?
• If I gave you your raw data today, could you reproduce this table?

If AI made any of these decisions for you, go back and make the decision again yourself.

Which tools should you use?

Use the tool your department accepts, and pair any chat assistant with software that shows its working. No single tool suits everyone. These details are as of November 2026, and they change often.

• jamovi is free and open source. It has a spreadsheet style and can show the R code behind each analysis, so you can learn as you click. The desktop version keeps data on your computer.
• JASP is free and open source, supported by the University of Amsterdam. It offers classical and Bayesian analyses and tables formatted for APA style.
• Google Colab is a free online notebook. Its Gemini Data Science Agent can build notebooks from plain-language goals. When Google opened it up in March 2025, it was for users aged 18 or over in select countries and languages, so check that it is available to you. Google itself warns that the agent may make mistakes. Your data goes to the cloud, so the zones above apply.
• Python in Excel needs a paid Microsoft 365 licence. Microsoft says free consumer and perpetual consumer licences are not supported.
• General chat assistants such as ChatGPT, Gemini, Claude and Copilot can explain methods and write code. Their free and paid plans differ, so check what your plan does with your data. Copilot Chat under a university work account may have different protections from a personal account. Ask your ICT office.

For finding and organising sources, see which AI tools work for literature reviews.

Questions people ask

Can I upload my survey data to ChatGPT or Gemini?

Only if your consent form, ethics approval and university rules allow it, and only after removing anything that identifies people. It is safer to give the tool a description and invented rows, then run the code yourself.

Is using AI for data analysis cheating?

It depends on your university's rules and on what the AI did. Getting help with code and a critique of your reasoning is very different from letting a tool choose your method and write your findings. Disclose what you did and ask your supervisor early.

Can AI choose the right statistical test for me?

It can suggest options and often does well on textbook cases. It may miss that your data breaks a test's conditions. Treat its suggestion as a starting point and confirm it against a statistics text and with your supervisor.

Do I still need to learn statistics?

Yes. You need enough to explain your method, spot a wrong result and answer a panel. AI can make learning faster, because you can ask it to explain a concept using your own variables.

Can I use AI to do my thematic analysis?

You can use it to test a codebook or to look for passages you may have missed, if your ethics approval allows it. Most qualitative methods rely on the researcher's own reading, so the themes should come from you. Say exactly what the tool did in your methods section.

What if my supervisor says no AI?

Follow that instruction. If you think a limited use would help, show your supervisor the fake-file method and ask about that one step. A written answer protects you both.

Do I need NACOSTI approval to use AI on my data?

NACOSTI's licence covers research in Kenya generally. We found no separate AI rule from NACOSTI. Read your licence conditions and ask your institution if they are unclear about outside data processors.

Sources

All links last checked 9 November 2026. Laws are summarised in plain English and this guide is not legal advice.

• Data Protection Act, No. 24 of 2019, Kenya Gazette copy on the ODPC website (sections 2, 49 and 53)
• NACOSTI, research licensing
• European Commission and ERA Forum, Living guidelines on the responsible use of generative AI in research, fourth version, 15 October 2026
• ICMJE, AI use by authors
• American Statistical Association, Statement on p-values (2016)
• Evaluating the Accuracy and Explanatory Quality of Large Language Models in Statistical Test Selection for Hypothesis Testing Decisions, Cureus (19 October 2025)
• Zhu et al., "Are Large Language Models Good Statisticians?", arXiv preprint (2024)
• Kirsten et al., "Decoding Complexity: Exploring Human-AI Concordance in Qualitative Coding", arXiv (2024)
• Koenecke et al., "Careless Whisper: Speech-to-Text Hallucination Harms", arXiv (2024)
• Ada Lovelace Institute, "Scribe and prejudice?" (11 February 2026)
• uni2study, "AI and ChatGPT Policies at Kenyan Universities" (7 August 2026)
• Google, Gemini Apps Privacy Hub
• jamovi
• JASP
• Google for Developers, Data Science Agent in Colab (3 March 2025)
• Microsoft Support, Python in Excel availability

Where FuKazee fits in

FuKazee runs the ScholarShift Accelerator for Masters and PhD students who want to use AI in their research without compromising academic integrity. It is a four-week programme with live online classes. It covers AI ethics in academia, prompting, literature reviews, citation management with Zotero, academic writing and defence preparation, and building custom research assistants. The next cohort is planned for February 2027.

We should be clear about what it is. It is not a statistics course, and it will not replace your supervisor or a statistician. It will help you work with AI carefully, keep your integrity and explain what you did.

We assume nobody knows AI yet, we avoid jargon, and we show the free version of each tool. We have trained more than 3,000 people.

To ask about the next cohort, email lead@fukazee.com.