What does responsible AI use look like for academic research, and how would a university know whether it is doing it well? Of thirty-eight top-tier doctoral universities surveyed across fifteen countries and jurisdictions, only six have AI policies that extend past research integrity into AI literacy and valid research practices. The world's major academic publishers issued a substantively identical no-AI-co-author policy across the sector within ten weeks of ChatGPT-3.5's public release. The United Kingdom's canonical PhD-researcher-development framework, refreshed in 2025, did not treat AI as a competency at all. While publishers and funders have responded to the emergence of AI, the universities that actually train researchers are lagging far behind what is needed to prepare them to use it responsibly. This report describes the significant opportunities and problems that agentic generative AI creates for research, setting out a competency framework for research training in the 21st century.
This report is written for senior university research leadership: Deputy Vice-Chancellors and Deputy Provosts of Research, Pro-Vice-Chancellors and Deputy Vice Presidents of Research, faculty Deans and Associate Deans of Research, Deans of Graduate Schools and Associate Deans of Research Training, higher-degrees and research-integrity committees, and the peak bodies that represent graduate students. Its purpose is to give that audience a shared, evidence-anchored vocabulary for benchmarking institutional readiness on responsible AI use in research and research training, and a maturity grid that translates the vocabulary into concrete decisions on policy, curriculum, infrastructure, and governance.
The report is the global anchor of a planned regional series. The same five-dimension competency framework and the same four-level maturity grid will be re-populated in later companion reports for the US / Americas, the UK / Europe, and APAC ex Japan.
This public repository contains the material needed to inspect the report's factual surface:
- the report source manuscript at
drafts/report.md; - the chart source and rendered chart assets in
charts/; - the web edition in
docs/; - the publication-formatted artefacts (DOCX, HTML, PDF) attached to the latest GitHub Release.
The report draws on eight primary-source evidence dossiers covering national research-funder AI policies (14 funders across 11 countries plus the EU's European Research Council), institutional AI policies at top-tier doctoral universities (38 universities across 15 countries and jurisdictions), existing AI-literacy and researcher-development competency frameworks (13 frameworks), publisher and journal AI policies (18 publishers plus 3 preprint servers), AI tool taxonomy (11 classes), privacy / IP / governance frameworks (14), the reproducibility-crisis intersection with AI augmentation, and adversarial-review patterns. The dossiers are working notes retained privately by Instats; every load-bearing claim in this report carries a primary-source URL with a 2026-05-18 snapshot date in the body footnotes.
Instats publishes the report source, the chart material, and the release files needed to inspect the report. Working files are retained privately by Instats. The author retains responsibility for every numerical claim, interpretation, and recommendation. Questions about public evidence can be sent to support@instats.org.
| Read online | https://mzyphur.github.io/responsible-ai-in-research-training/ |
| Microsoft Word (.docx) | direct download: https://mzyphur.github.io/responsible-ai-in-research-training/Instats%20-%20Responsible%20AI%20in%20Academic%20Research.docx |
| direct download: https://mzyphur.github.io/responsible-ai-in-research-training/Instats%20-%20Responsible%20AI%20in%20Academic%20Research.pdf | |
| HTML | direct download: https://mzyphur.github.io/responsible-ai-in-research-training/ |
| Releases page (all versions) | github.com/mzyphur/responsible-ai-in-research-training/releases |
| Markdown source | drafts/report.md |
Downloading. Use the direct-download links above (which serve the actual file with the correct
Content-Type) or the release-page assets. If you click into thedocs/folder via GitHub's file tree and "Save Link As" on the file there, GitHub serves the browser preview page — not the file — and the saved file will not open in Word. Use the direct-download links above instead.
Citation. Zyphur, M. J. (2026). Responsible AI in Academic Research: A Competency Framework for Research Training. Instats Policy Series. https://github.com/mzyphur/responsible-ai-in-research-training. ORCID: 0000-0003-3237-7892. DOI: 10.61700/t31oy23grr.
Available as HTML (read online), PDF (direct download), Word (direct download), and source repository (GitHub).
BibTeX:
@techreport{zyphur2026responsibleai,
author = {Zyphur, Michael J.},
title = {Responsible AI in Academic Research: A Competency Framework for Research Training},
institution = {Instats},
type = {Instats Policy Series},
year = {2026},
url = {https://github.com/mzyphur/responsible-ai-in-research-training},
note = {ORCID: 0000-0003-3237-7892. DOI: 10.61700/t31oy23grr.},
doi = {10.61700/t31oy23grr}
}Machine-readable citation: CITATION.cff.
The framework's spine, set out in Part 2 of the report and operationalised as a four-level maturity grid in Part 4 and Appendix A:
- Human-in-the-loop discipline — the institutional commitment that the judgement steps which define research remain human, with task-level demarcation between labour and judgement.
- Responsible use in practice — operational rules for the four AI-use modes (search; co-author; validator; tutor) at PhD level.
- Tooling that promotes responsible use — institutional procurement standard incorporating six observable properties (verifiable citation, data residency, uncertainty reporting, reproducibility, auditability, open-source-and-local options).
- AI-literate humans — six load-bearing competencies for PhD researchers, supervisors, and examiners (citation verification; model-and-parameter specification; prompt-as-fork-in-the-garden discipline; model-heterogeneity in adversarial review; sycophancy detection and human-as-verifier discipline; structured failure-mode reporting).
- Institutional benchmarking grid — the institutional competency to know where it sits on Dimensions 1-4 and act on what the answer reveals, scored across four axes: policy / people / systems / process.
Appendix G of the report carries a worked labour-vs-judgement task taxonomy that institutions can adopt and adapt.
responsible-ai-in-research-training/
├── docs/ ← published GitHub Pages site (HTML + PDF mirror)
├── drafts/
│ └── report.md ← markdown source manuscript
├── final/
│ ├── report.docx ← Microsoft Word build (Word-for-Mac compatible)
│ └── reference.docx ← Word styling template
├── charts/
│ ├── 01_regulatory_response_timeline.py ← Figure 1 source
│ ├── png/ · svg/ ← rendered chart assets
│ └── style.py ← chart stylesheet
├── assets/
│ └── instats_logo.png
├── CITATION.cff ← machine-readable citation (CFF 1.2)
├── LICENSE ← CC BY-NC-ND 4.0
├── README.md ← this file
├── SECURITY.md
└── VERSION ← single source of truth (1.1.0)
Working notes (evidence dossiers, review files, and launch material) are retained privately by Instats and are not part of this public repository.
Before release, the manuscript went through several rounds of independent review: adversarial fact-checking against the primary sources, a reader-persona panel spanning the five audience tiers, multi-pass copy-editing, and a final read for voice and clarity. Findings from those rounds were resolved across the v0.5.x and v0.6.0 revisions and closed out at v1.0.0.
The author used author-directed computational and research-assistance workflows while preparing this report; Appendix D of the report carries the full AI-assistance disclosure. The author retains responsibility for every numerical claim, interpretation, and recommendation, and every load-bearing claim was verified against the primary sources by the lead author before publication.
|
Michael J. Zyphur, PhD Instats | instats.org support@instats.org |
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) — see LICENSE. © 2026 Instats. Share in full, with attribution, for non-commercial purposes; no adapted or modified redistribution; commercial reuse by written permission from Instats.
Instats Policy Series · 2026 · Published openly so any reader can audit the evidence and the framework.