Start with one safe research task

You do not need to know PISA or topological data analysis. You do not need to complete a stack of forms or learn a specialised prompt formula before trying generative AI (GenAI).

Three steps for a first encounter

1. Choose a safe, bounded task. Begin with a generic, non-confidential question from your research. For example, ask what different comparisons might reveal in a proposed survey study, or what rival readings of a published policy statement you should consider. Check permissions before sending material: even a summary or an extract can disclose restricted information. When uncertain, use a wholly invented example and share no research records.

2. Ask for possibilities, then check one. Ask the assistant for alternatives, assumptions and reasons its suggestions might fail. The answer is a proposal, not an executed analysis, checked reference or finding. See the constructed first GenAI exchange, which includes a prompt you can adapt using permitted material.

3. Record only a consequential decision. If a proposal actually changes your question, design, interpretation or writing, briefly record the proposal, your choice and why you made it. Keep that initial entry separate from later updates. The Decision Log is optional; ordinary dated research notes may be sufficient.

For the six practical safeguards, consult the field guide. You can stop there for an exploratory session.

Go further when a problem arises

• AI suggests code or an analysis: Generated analysis is not evidence.
• A number appears right but its interpretation is wrong: Plausible number, wrong inference.
• A preferred method meets inconvenient evidence: Evidence narrows the method claim.
• Someone claims AI originated an idea: Retrospective storytelling.

These PISA examples are bounded accounts of a real research programme, not proof that the framework works or a full public record of the project. The research decision guide helps with other questions. Use the ten-step protocol and crosswalk as references when a project becomes more complex, not as prerequisites.

Researchers working with interviews, archives, documents or coded corpora should adapt checks to their own sources, methods and ethics. Statistical replication is not the only form of research verification. For guided group discussion, see the teaching activities.