QuantLab · MDAP · CAIR
AI in
Research
Brown Bag workshop series
Session 1 — Getting started with AI assistance for research
Monday 10 August 2026 · 12–1pm, drop-in until 1.30pm
Manhari Room, Melbourne Connect
Acknowledgement of Country
The University of Melbourne acknowledges the Traditional Owners of the unceded land on which we work, learn and live: the Wurundjeri Woi-wurrung and Bunurong peoples (Burnley, Fishermans Bend, Parkville, Southbank and Werribee campuses), the Yorta Yorta Nation (Dookie and Shepparton campuses), and the Dja Dja Wurrung people (Creswick campus).
The University also acknowledges and is grateful to the Traditional Owners, Elders and Knowledge Holders of all Indigenous nations and clans who have been instrumental in our reconciliation journey.
We recognise the unique place held by Aboriginal and Torres Strait Islander peoples as the original owners and custodians of the lands and waterways across the Australian continent, with histories of continuous connection dating back more than 60,000 years. We also acknowledge their enduring cultural practices of caring for Country.
We pay respect to Elders past, present and future, and acknowledge the importance of Indigenous knowledge in the Academy. As a community of researchers, teachers, professional staff and students we are privileged to work and learn every day with Indigenous colleagues and partners.
In making this Acknowledgment of Country we commit to respectful and responsible conduct towards all others according to the Traditional lores of this land, particularly at times of formal ceremony.
These monthly lunchtime sessions are a relaxed, informal space to learn, experiment, and connect with others across the University who are working out how to use AI in their research.
Open to researchers at all levels and all disciplines. No prior experience with AI assumed, and nothing here is compulsory — if you only want to watch, that’s fine.
Brought to you by QuantLab, the Melbourne Data Analytics Platform (MDAP), and the Community of AI in Research (CAIR).
Every session gives you three things
12.00 – 12.20
Intro
A speaker introduces a use case of AI in research. Questions welcome throughout.
12.20 – 1.00
Hands-on
You follow along on your own laptop. We circulate and help. Nothing to prepare beyond the setup we send in advance.
1.00 – 1.30
Optional drop-in
Stay longer if you like. Bring your own project, your own data, or a question which has not been addressed.
All sessions run at Melbourne Connect, 12–1pm, with the drop-in until 1.30pm. Rooms vary — check the flyer or your registration email.
12.00
Welcome — launching the series Seraphine Maerz & Mel Mistica
12.05
Research integrity and AI Mia Jackson, OREI
12.10
Copyright and AI Sharon Stewart, Copyright Office
12.15
Questions for Mia and Sharon
12.25
AI in research: a brief introduction Seraphine
12.35
Claude Code — an AI agent that works on your files Laptops out, Seraphine and Kim (online)
1.00
Optional drop-in — bring any AI question Mia, Seraphine
Over to our guests
Research integrity,
copyright and AI
The rules you’re actually working under: disclosure, authorship, data handling, copyright, and the new University’s policy.
Guest speaker · 5 minutes
Mia Jackson
OREI
Guest speaker · 5 minutes
Sharon Stewart
Copyright Office
10 minutes
Questions for Mia and Sharon
Anything about integrity, ethics, copyright, disclosure or data handling.
Part two
AI in research
A very quick overview.
AI as research assistance
It helps you do the work: cleaning data, writing scripts, converting files, drafting, explaining. The model is nowhere in your findings.
What you report: a detailed disclosure statement in line with university policies — which tool, which version, what it did.
AI as research method
The model generates your data: coding, classification, annotation, extraction. The model is your instrument, the way a survey item or a codebook is.
What you report: model, version, date, parameters, the full prompt, all replication scripts, and validation against human coding.
Do it with quallmer (R): codebooks, Krippendorff’s α, audit trails. quallmer.github.io/quallmer · quantilab.github.io
If it is method, it belongs in your research design from the start — chosen, justified and validated like any other instrument.
Developed by Anthropic with Rick Dakan (Ringling College) and Joseph Feller (University College Cork). Four competencies, easy to remember, and the middle two are where most research goes wrong.
1 · Delegation
Deciding what to hand over and what to keep. Which parts of this task genuinely need my judgement?
2 · Description
Communicating what you actually want — context, constraints, and what “finished” looks like. Not clever phrasing.
3 · Discernment
Evaluating what comes back: is it accurate, is it sound, is the reasoning real? Critical thinking cannot be delegated.
4 · Diligence
Owning the result. Being transparent about how it was produced, and accountable for it either way.
Privacy and data security
Is this data allowed to leave the University? Is the tool allowed to be used (check UniMelb IT in case of doubts!)
Copyright
Whose work goes in — and whose comes out?
Bias & fairness
Models carry the biases of their training data. Who could that harm?
Accountability
You are responsible, not the model. Disclose the use.
Transparency
For closed models, the training data is simply unknown.
Reproducibility
The same prompt can give a different answer tomorrow.
Resource use
Large models cost real energy and water.
Remember
Open-weight models for sensitive data. Traceability for every result. Always validate.
Closed & hosted (Claude, ChatGPT, Gemini): someone else runs the model — your data goes to them, the model version can change anytime. Open-weight (Llama, Qwen, Mistral): you download the model and run it yourself, so nothing leaves, and you can freeze a version for better reproducibility.
Chat assistants
You ask, it answers.
Claude · ChatGPT · Gemini · Copilot
Agentic tools
They don’t just answer — they act on your files and run things. Today’s topic.
Claude Code · Claude Cowork · OpenAI Codex · Gemini CLI · Cursor · GitHub Copilot
Run it yourself
Open-weight models on your own machine or on University infrastructure - better option for research!
Ollama · BERT models · vLLM · Hugging Face · OpenCode
Literature and discovery
Might be useful for finding and triaging — does not replace reading the paper.
Elicit · Semantic Scholar · Gemini Notebook
Assume this slide is already out of date.
Part three · 25 minutes
Claude Code
One agentic tool, the way it works generalises (more or less) to the rest.
Working with this (fee-based) tool is optional!
…and the hands-on exercises are optional!
You can also just sit back and watch, if you prefer!
Chat
You copy something in. It answers. You copy the answer out and find out for yourself whether it was right.
It can’t see your files, can’t run your code, and never learns whether its suggestion worked.
Agent
It works inside a folder on your machine. It reads your files, writes files, runs commands, reads the errors that come back, and tries again.
You approve each action. You see every change before it happens.
The loop: you ask → it reads → it acts → it checks whether that worked → it corrects itself → you verify. The self-correction is what makes the output better. Your verification is what makes it usable.
It isn’t a chatbot that writes code. It’s a very fast, very literal colleague working in your folder — one who needs supervising!
Not on this list: writing your paper. The value is in the plumbing around the research, not the thinking that is the research.
1 · Check IT compliance and data sharing rules
Observe (changing!) UniMelb IT policies. On a personal Free, Pro or Max account Anthropic may train on your prompts, and keeps a copy of them for five years. Switch that off at claude.ai/settings/data-privacy-controls: no training, and copies are deleted after 30 days.
2 · Open it in a project folder, never your whole computer
It sees the folder you point it at and everything below — nothing above. Pick the project folder, not your home folder or Desktop. What it cannot reach, it cannot read, send or change. Put in this folder only what you can share.
3 · Write a CLAUDE.md
/init drafts one. Re-read every session: what the project is, which folders are off limits, what to ask you before doing. Clear usage rules and control.
4 · Put the off-limits folders in writing
A .claude/settings.json file lists what it may never open (“deny” rule). Adding there Read(./data/raw/**) blocks everything in the raw-data folder, for example. A deny always beats a permission, and the .json file travels with the project.
My advise: Use the app, not the terminal. On University-managed macs, for example, the terminal can usually reach your whole disk — and Claude Code would inherit that same access. The app can only reach the folder you open in the app.
If you do use the terminal, sandbox (=isolate) it — see code.claude.com/docs/en/sandbox-environments
1 · Install
Desktop app — macOS or Windows. Download and install from claude.ai/download (not a search result), then open it and sign in with your Claude account.
What you need to have set up
A Claude subscription such as Pro. Free accounts can’t run Claude Code. If that’s you or if you do not want to sign up: pair up with the person next to you.
2 · Point it at a folder
Create a new, empty folder called brownbag somewhere you can find it. In the app, click Code → Local → Select folder and point it at that folder — nowhere else.
Select folder: choose the project folder only — never your whole Desktop or home folder or any other folder you do not want to share!
3 · Switch to Plan mode
Set the mode selector next to the send box to Plan — it proposes before it acts. Start here every time.
Type each boxed prompt below into the app — your own wording works just as well, this is only a starting point.
1 — Make some material
Invent 20 survey responses on AI tools: rating 1–5, comment on some. Show a few, save as .csv.
2 — Analyse it, and show it
Build and save one HTML page: ratings chart, plus a clickable card per identified theme. Open it.
3 — Check it
Pick two comments. Why did you group them that way?
4 — Make it transparent
Add a note at the top: who you are, what you did, how, what to check. Open it again.
Watch the middle bit — what happens between your prompts, not the answer. Esc interrupts at any time.
Validate everything
Nothing an AI produces is evidence until you have checked it against something that is.
Agents act
That’s what makes them worth the trouble, and it’s what changes the risk profile. Supervision is the price of the capability.
Disclose it, commit it
A disclosure sentence naming the tool, and the artefact under version control.
The 4Ds are a good pocket checklist: delegate deliberately, describe properly, discern — never skip this one — and be diligent about saying what you did.
Any questions?
Mon 14 September · Manhari Room
Equipping AI agents with tools
Daniel Russo-Batterham, MDAP
Mon 12 October · Room M13
Working with synthetic data
Irma Mooi-Reci, School of Social and Political Sciences and QuantLab
Mon 9 November · Room M01
Making AI more trustworthy with knowledge maps
Taylor Ey, Melbourne Institute
Mon 14 December · Manhari Room
Analysing historical data with AI
David Goodman, School of Historical and Philosophical Studies
All sessions 12–1pm at Melbourne Connect, with the optional drop-in until 1.30pm.
Optional
Stay for the drop-in
Until 1.30pm. Any AI question at all — bring your own project, your own data, or the thing you’ve been meaning to ask someone about for six months.
These slides are available here quantilab.github.io/sharezone/
Most of these slides were created with the assistance of Claude Code.
Thank you — see you next month.
QuantLab · MDAP · CAIR