AI in Research: Methods and Policy

Research methods and policy for using generative AI while keeping evidence, reliability, and accountability with the researcher.

How can researchers use AI in work that others can check? We examine how to keep a clear record of what the AI did and which evidence supports the result. The researcher remains responsible for every decision.

One part of this work turns policy into decisions researchers can act on. Our university guidance sets boundaries for appropriate AI use and explains what researchers need to disclose. Our research on institutional frameworks compares how universities developed guidance that permits useful work while keeping human oversight.

We also build and document methods. Our document-discovery work uses language models to classify archived web pages. Each classification retains a supporting quotation, and the packaged process lets another researcher inspect or repeat it. A related methodological study asks what researchers must design around a model to make AI-assisted work reliable. Together, these projects connect practical tools with clear responsibility for their use.

Ballsun-Stanton, B. (2026, August 11). The Ethics of Investigating Digital Worlds [Lecture]. INTS1302 Navigating Digital Worlds, Macquarie University. https://techethicsai.au/course/INTS1302-week03/week03.html#/

Presentation

Ross, S., & Ballsun-Stanton, B. (2026, July 30). Paper B: Reliability in research with large language models is a property of the human–AI system. OSF. https://doi.org/10.17605/OSF.IO/M376W

Preprint

Why it matters

The real question is not whether to use these systems but how to deploy them so that research tasks can be completed and their output can be trusted. Verification is where practice needs development.

Source: p. 3

Read abstract

Companion paper B (submitted to Royal Society Open Science, which uses open peer review): a methodological case study of large language model (LLM)-assisted research, arguing that reliability must be engineered into the human–AI system rather than expected of models alone. This component holds the preprint, electronic supplementary material, and the AB+ literature-grounding pipeline and outputs; shared data live in the Software Tools Dataset and LLM Research-Assistance Corpus components.

Ballsun-Stanton, B., Johnston, V. H., Jensen, H. S., Kjeldsen, C. K., & Thøgersen, J. (2026, April 15). LLM Document Discovery (Version Container) [Computer software]. Zenodo. https://doi.org/10.5281/ZENODO.19591245

Software

Why it matters

Our method lets historians and other researchers work from a problem-oriented approach where the research questions lead the search for relevant source material.

Source: article notebook, Markdown cell 8

Read abstract

LLM Document Discovery v0.2.0 — Apptainer Container Pipeline

Reproducible pipeline for classifying historical web documents (1996–2005) using large language models. Extracts linguistic and structural features from children's web

content archived by the Internet Archive, producing a structured SQLite database of classifications with supporting blockquote evidence.

This release adds a containerised execution pipeline using Apptainer/Singularity, enabling reproducible deployment on both local GPUs and HPC clusters (NCI Gadi).

What's new in v0.2.0:

Apptainer container: wrapping vLLM v0.19.0 + llm-discovery for fully offline, reproducible execution on HPC compute nodes

Single-command deployment: deploy assembles the data directory, syncs to HPC, and submits the PBS job

Multi-model support: tested with google/gemma-4-E4B-it (local RTX 4090), google/gemma-4-31B-it (Gadi V100), and openai/gpt-oss-120b (Gadi H200)

Prompt engineering: rationale-first instruction format achieving 98% structured output compliance (up from <1%)

Crash-safe resumability: killed containers resume from the last completed document–category pair

CLI commands: build, init, download-model, deploy, status –watch, retrieve, for end-to-end HPC lifecycle management

Offline tokeniser support: tiktoken vocab files baked into the container for OpenAI gpt-oss model family

Associated paper:

Johnston, V.H., Ballsun-Stanton, B., Jensen, H.S., Kjelsen, C.K., & Thøgersen, J. (2026). Exploring the Archived Web through AI-Assisted Document Discovery. https://github.com/WEB-CHILD/exploring-the-archived-web-through-ai-assisted-document-discovery

Source notebook

Johnston, V. H., Ballsun-Stanton, B., Jensen, H. S., Kjeldsen, C. K., & Thøgersen, J. (2026, April 13). Exploring the Archived Web through AI-Assisted Document Discovery. Journal of Digital History. https://journalofdigitalhistory.org/en/notebook-viewer/JTJGcHJveHktZ2l0aHVidXNlcmNvbnRlbnQlMkZXRUItQ0hJTEQlMkZleHBsb3JpbmctdGhlLWFyY2hpdmVkLXdlYi10aHJvdWdoLWFpLWFzc2lzdGVkLWRvY3VtZW50LWRpc2NvdmVyeSUyRm1haW4lMkZhcnRpY2xlLmlweW5i/?v=3

Paper

Why it matters

By categorising each source from the Kidlink domain into our categories we have been able to find material that we did not know existed on the domain.

Source: Results of last iteration

Source notebook

Smith, S. M., Tate, M., Freeman, K., Walsh, A., Ballsun-Stanton, B., & Lane, M. (2026, January 2). A university framework for the responsible use of generative AI in research. Journal of Higher Education Policy and Management, 48(1), 17–36. https://doi.org/10.1080/1360080X.2025.2509187

Paper

Why it matters

In 2023, Macquarie University and Queensland University of Technology each consulted researchers to develop a position on AI use in research. Those positions allow researchers to use AI where it helps while requiring them to disclose its use and remain accountable. We draw on both cases to offer a four-step process that other universities can adapt.

Ballsun-Stanton, B., & Smith, S. (2025, April 30). Generative AI in Research: Disclosure & Documentation Guidance / Checklist. Macquarie University. https://doi.org/10.5281/ZENODO.15307492

Guidance

Why it matters

Keep dated records of your prompts, outputs and later edits so other researchers can verify what you did. Match the disclosure to how extensively AI contributed.

Read abstract

The use of generative AI in research needs to align with the principles and responsibilities outlined in the Australian Code for the Responsible Conduct of Research 2018. Guidelines about the appropriate disclosure and documentation to accompany the use of Generative AI in research also need to recognise the practical considerations of current or innovative research practices. An approach combining these elements is detailed in this document for Macquarie University researchers.

Ballsun-Stanton, B. (2025, February 18). Research Use of LLMs: Beyond the chatbot. https://doi.org/10.5281/ZENODO.14885316

Presentation

Why it matters

Move beyond a chat window when research needs to process many sources or make AI-assisted work reproducible. Connecting through code gives tighter control over data, prompts and model settings.

Read abstract

Slides for the Research Data Alliance event:

Title: AI in Action: How Researchers Leverage AI (Asia/Oceania Friendly Time)

https://www.rd-alliance.org/event/ai-in-action-how-researchers-leverage-ai-europe-americas-friendly-time-2/

Ballsun-Stanton, B. (2024, July 19). Avoiding Sadness: Research policy for Generative AI [Keynote]. https://denubis.github.io/germany-keynote-ai-policy-briefing/avoiding-sadness

Presentation

Why it matters

Research policy can allow generative AI while setting narrow limits where human judgement cannot be delegated. Researchers remain accountable for how they use it.

Recording of the research methods guest lecture version

Ballsun-Stanton, B. (2024, February 12). Briefing on the Guidance Note on Generative AI in Research at Macquarie University. https://doi.org/10.5281/ZENODO.10648834

Presentation

Read abstract

A briefing on the Macquarie University Generative AI in Research Guidance Note. The Guidance note can be found at https://policies.mq.edu.au/download.php?associated=1&id=768&version=1

Ballsun-Stanton, B., Smith, S. M., & White, K. (with Anson, D., Connor, M., Cornish, J., D’Arcens, L., Kim, J., Singh, A., & Todd, V.). (2023, October 22). Guidance Note: Using Generative Artificial Intelligence in Research. Macquarie University. https://doi.org/10.5281/ZENODO.10851623

Guidance

Read abstract

This is the university policy/guidance note for the use of Generative AI and LLMs at Macquarie University, Australia. The original source is https://policies.mq.edu.au/download.php?associated=1&id=768&version=1

Here is the summary:

Generative AI (Artificial Intelligence, Large Language Models) offers the unprecedented ability tomanipulate and generate text and media in response to arbitrary instructions. These new capabilities offeropportunities and risks to researchers. This document will discuss responsible use, risk mitigation, andappropriate use of these tools. The technological landscape changes quickly and new tools are releasedalmost weekly – this guide offers general advice which should be applied thoughtfully.

Regardless of any tools or technologies used, now or in the future, everyone at Macquarie University isresponsible for ensuring their research meets the expectations of the Australian Code for the ResponsibleConduct of Research and the Macquarie University Code for the Responsible Conduct of Research(2018).

Generative AI must be used with caution, and its use is currently inappropriate in some researchprocesses because Generative AI services, including ChatGPT:• cannot meet the requirements for authorship• can create authoritative-sounding outputs that may be incorrect, incomplete, or biased• could inappropriately capture sensitive data (including, but not limited, to personal information).

Researchers must not use Generative AI:• to perform peer review activities• to generate substantive content of research outputs, including HDR theses• for writing the critical components of human ethics, animal ethics, or biosafety applications.

Researchers must exercise care in the use of Generative AI in other aspects of their research and should:i. only do so with the written agreement of their research collaborators (& HDR supervisors)ii. review and consider the terms of service/license of the platforms used and any models usediii. consider the current issues and understandings around copyright and intellectual propertyiv. mitigate risks around the insecure storage or unauthorised re-use of sensitive datav. exert oversight and control when using the technologyvi. carefully and critically review the output and results created by Generative AIvii. take responsibility for the integrity of the content altered or created using Generative AIviii. disclose the use of Generative AI to potential publishers and in disseminated research outputsix. read and follow the policies of publishers and funders regarding the use of Generative AI.