1. Whenever possible, use open models. If not possible, appropriate the capabilities of the proprietary model by training your model with it.

  • Fine-tuned open models match or exceed GPT-41

  • LoRA and RAG enable corpus-grounded, domain-adapted analysis2

  • Distillation via synthetic data legitimately transfers proprietary capabilities3

  • Smaller models sometimes outperform their teacher model4

  • Proprietary models in research require explicit justification5

2. Whenever possible, use your own, sustainable infrastructure.

  • Local hosting preserves data, technological, and infrastructural sovereignty6

  • Proprietary tools routinely fail data sovereignty obligations7

  • Inference is now AI’s dominant lifecycle environmental cost8

  • Quantisation and local inference cut emissions ~45%9

  • Institutional computing complements rather than duplicates national infrastructure10

3. Never, ever, compromise the privacy of your research partners by uploading your data to a “free-to-use” LLM API.

  • “Free-to-use” means paid for in data alienation: by default, AI companies train on what users type11 12

  • Commercial APIs transfer participant data to external servers13

  • There is no reliable way to make models forget14

  • Research data extraction continues colonial histories of dispossession15

  1. Carammia, M., Iacus, S. M., & Porro, G. (2024). Rethinking Scale: The Efficacy of Fine-Tuned Open-Source LLMs in Large-Scale Reproducible Social Science Research. https://arxiv.org/abs/2411.00890 ↩︎

  2. Ngo, T. T., Van, D. N., Nguyen, A.-M., Do, P.-A., & Nguyen-Quoc, A. (2026). Qualitative Coding Analysis through Open-Source Large Language Models: A User Study and Design Recommendations. https://arxiv.org/abs/2602.18352 ↩︎

  3. He, B., et al. (2026). Pedagogically-Inspired Data Synthesis for Language Model Knowledge Distillation. https://arxiv.org/abs/2602.12172 ↩︎

  4. Woo, G. W., et al. (2024). Synthetic Data Distillation Enables the Extraction of Clinical Information at Scale. https://www.medrxiv.org/content/10.1101/2024.09.27.24314517 ↩︎

  5. Palmer, A., Smith, N. A., & Spirling, A. (2024). Using proprietary language models in academic research requires explicit justification. Nature Computational Science, 4(1), 2–3. https://doi.org/10.1038/s43588-023-00585-1 ↩︎

  6. Kaffee, L.-A. & Jernite, Y. (2025). Open Source AI: A Cornerstone of Digital Sovereignty. Hugging Face. https://huggingface.co/blog/frimelle/sovereignty-and-open-source ↩︎

  7. García-Verdugo, H. D. & Román-Palacios, C. (2025). LabOps: A flexible self-hosted workflow of open source tools for efficient collaboration within research laboratories. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12212491/ ↩︎

  8. Jegham, N., Abdelatti, M., Elmoubarki, L., & Hendawi, A. (2025). How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference. https://arxiv.org/abs/2505.09598 ↩︎

  9. Khan, T., et al. (2025). Optimizing Large Language Models: Metrics, Energy Efficiency, and Case Study Insights. https://arxiv.org/abs/2504.06307 ↩︎

  10. Kitaeff, S., et al. (2025). Institutional Research Computing Capabilities in Australia: 2024. https://arxiv.org/abs/2509.17351 ↩︎

  11. Couldry, N., & Mejias, U. A. (2019). Data Colonialism: Rethinking Big Data’s Relation to the Contemporary Subject. Television & New Media, 20(4), 336–349. https://journals.sagepub.com/doi/10.1177/1527476418796632 ↩︎

  12. King, J., et al. (2025). User Privacy and Large Language Models: An Analysis of Frontier Developers’ Privacy Policies. https://arxiv.org/abs/2509.05382 ↩︎

  13. Ngo, T. T., Van, D. N., Nguyen, A.-M., Do, P.-A., & Nguyen-Quoc, A. (2026). Qualitative Coding Analysis through Open-Source Large Language Models: A User Study and Design Recommendations. https://arxiv.org/abs/2602.18352 ↩︎

  14. Yu, J., He, Y., Goyal, A., & Arora, S. (2025). On the Impossibility of Retrain Equivalence in Machine Unlearning. https://arxiv.org/abs/2510.16629. See also Lynch, A., Guo, P., Ewart, A., Casper, S., & Hadfield-Menell, D. (2024). Eight Methods to Evaluate Robust Unlearning in LLMs. https://arxiv.org/abs/2402.16835 — empirically demonstrating that current unlearning methods fail under adversarial probing. ↩︎

  15. Bridge, E. & Block, C. (2025). AI Reflections: Indigenous Data Sovereignty and Artificial Intelligence. University of British Columbia. https://indigenousinitiatives.ctlt.ubc.ca/2025/11/19/ai-reflections-indigenous-data-sovereignty-and-artificial-intelligence/ ↩︎