# Use AI to build tools while retaining scientific judgement

Generative AI assistants can support programming, tool development and iterative revision. I distinguish executable code, a valid statistical analysis and an answer to the right research question. Specify the estimand, population, timing and missing-data assumptions before interpreting a result. Human responsibility includes confidentiality, equity and the ability to challenge an output.

## 1. Frame

Purpose, permitted data and expected behaviour.

## 2. Build

Small, understandable changes.

## 3. Challenge

Inspect code, calculations and assumptions separately.

## 4. Account

Document contributions and human decisions.

## Fictional situation · proposed approach

A tool must explain a research finding to managers and the public. The level of delegation depends on whether the task concerns code, analysis or learning.

### Develop a tool

Reasoning : Specify inputs, outputs and edge cases. Review the changes and use synthetic data when appropriate.

Useful output : A specification, version history and an independently calculated reference example.

What to watch : An agent’s access to files and services must fit the authorised task.

### Analyse and interpret

Reasoning : Write the question and assumptions first. Request alternatives and distinguish planned analysis from exploration.

Useful output : An analysis log covering population, variables, estimand, missing data and sensitivity analyses.

What to watch : Producing more can create an illusion of understanding. Preserve methodological diversity and contestability.

### Learn with PAIRR

Reasoning : Start personally, discuss with a peer, use AI as a second critique and justify accepted and rejected suggestions.

Useful output : Successive versions, a reflective note and an appropriate transfer task without assistance.

What to watch : The quality of the final product does not prove learning. State access and assessment rules explicitly.

## Questions to consider

- An agent’s access to files and services must fit the authorised task.
- Producing more can create an illusion of understanding. Preserve methodological diversity and contestability.
- The quality of the final product does not prove learning. State access and assessment rules explicitly.

## Sources and scope

Experimental results, conceptual arguments and governance guidance answer different questions. They inform decisions without supporting universal effectiveness claims. Documentary cut-off: 15 October 2026.

Messeri, L., & Crockett, M. J. (2024). Artificial intelligence and illusions of understanding in scientific research. Nature, 627, 49–58. https://doi.org/10.1038/s41586-024-07146-0

Car, J., Ke, Y., Jin, L., Ong, J. C. L., Thirunavukarasu, A. J., Cheung, C. Y., Tham, Y. C., Ting, D. S. W., Ong, M. E. H., Compton, S., Narayan, A., Keane, P. A., Wong, T. Y., Bates, D. W., Tan, P., & Liu, N. (2026). AI-induced never-skilling in medical education. Nature Medicine, 32, 1997–2006. https://doi.org/10.1038/s41591-026-04438-y

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122

Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15(1), Article 17458. https://doi.org/10.1038/s41598-025-97652-6

World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. https://www.who.int/publications/b/70584