Choose a question you can actually investigate
The most useful student project is not always the largest. A small question with a clear method can teach more than a broad application whose behaviour nobody can explain. Start in a setting you understand: a library, a classroom exercise or a public-information page. Ask what someone needs to know and where confusion occurs. Avoid making decisions about real applicants, patients or classmates. For independent practice, use public material you are permitted to reuse or create synthetic examples that do not describe identifiable people.
Keep the source close to the claim
Suppose you are exploring whether an assistant can summarise a public guide. Save the exact version you used, its source and the date you accessed it. Then compare the summary with that reference sentence by sentence. Separate what the source explicitly says from what the system inferred. A fluent explanation is not evidence of accuracy. If a reader would need another document to verify a claim, identify that gap rather than silently treating the claim as established. A modest, well-supported summary is stronger work than a comprehensive-sounding invention.
Make a test another person can repeat
Write down the task, input, instructions and review criteria before you inspect the output. Include at least one ordinary example, one ambiguous example and one example that should produce an admission of uncertainty. Keep enough context for a peer to repeat your method, while respecting licence terms and data restrictions. You do not need to publish credentials, private conversations or internal documents. If reproducibility would require sharing restricted information, redesign the exercise around synthetic material instead.
Treat the failure as the interesting part
Imagine the system correctly summarises opening hours but invents an appointment requirement. Do not hide that answer and replace it with the best result. Explain what was invented, why it matters and how your review caught it. Then change one part of the exercise and test again. Perhaps you clarify that every factual claim needs support from the reference, or add an explicit instruction to acknowledge missing information. Report whether the change helped on new examples, not only on the case that prompted it. This keeps the lesson honest.
Ask for a specific kind of feedback
“What do you think?” invites encouragement but often misses the methodological weakness. Ask a peer to check whether your acceptance criteria match the task, whether a claim lacks support or whether a non-AI approach would be simpler. If you involve a teacher or professional, agree a bounded request rather than assuming an ongoing mentoring relationship. A careful resource review can be a contribution in its own right. It does not need to become a product, a placement or an institutional partnership to be worthwhile.
Use a field-note structure
Open with the question and why it matters. Describe the reference material and permissions. Explain the method, show a representative result and a failure, and state what you changed. Finish with limitations and the next question. Credit contributors only with their consent, and distinguish an individual contribution from institutional endorsement. A useful note can be short, but it should let someone understand what happened without watching a staged demo. If the result was inconclusive, say what evidence you would need to decide.
Your next small contribution
Choose one public reference and write three synthetic questions about it. Identify in advance what an acceptable answer must include and what it must not invent. Review the outputs with a peer and write a field note from the disagreement. The Council’s proposed learning route would build on this kind of practice; mentors, safeguarding and participation terms are not yet confirmed. This article offers an exercise you can discuss, not an application or a promise of recognition.
Sources & editorial context
Original discussion analysis, not a reported interview or programme result. Synthetic examples are illustrative. Related reading provides context, not endorsement.
BPOAI — The Team Behind BPOAI.ai Joins NVIDIA Inception (9 October 2026) ↗
POEA Job Finder — Andy Schachtel Keeps People in the Loop (9 October 2026) ↗