What you will learn
  • Turn a broad topic into a researchable question.
  • Distinguish primary evidence from commentary.
  • Check that a citation supports the attached claim.
  • Record uncertainty and conflicting evidence.
  • Write a conclusion whose strength matches its sources.

Before you begin

Read USE03 for task briefs and USE04 for focused questions. You need a browser and a small research topic; an AI tool with live browsing is optional.

Begin with a question that evidence can answer

‘Research AI robots’ is too broad to guide a useful search. A narrower question is ‘What information would I need to decide whether a classroom robot should process camera images locally?’ It identifies a decision and exposes subquestions about response time, hardware, privacy, and reliability. The assistant can help organize those questions before making claims.

Write the scope: who the answer concerns, what time period matters, and which decision it will support. A historical explanation of a technique needs different sources from a current compatibility check. Include what would change your mind. If you are comparing two options, define criteria before reading persuasive descriptions of either one.

Use the assistant to propose search terms and identify unfamiliar vocabulary. Treat the results as leads, not evidence. A generated bibliography can contain nonexistent titles, incorrect authors, or genuine papers that do not support the accompanying claim. Research begins when you inspect the sources.

Find the source closest to the claim

A primary source provides direct evidence for the question at hand: an original research paper, an official board datasheet, a released specification, or a dataset’s documentation. A tutorial or news report may explain the material well, but it is another layer between you and the evidence. Use it for orientation and follow its references where appropriate.

Primary does not mean infallible or unbiased. A manufacturer’s page is strong evidence for its documented connector layout but may be a weak basis for declaring its product universally best. A research paper reports a particular method, sample, and measurement. Its conclusion may not transfer to a different task or population.

Check the publication date and the date of the underlying event or experiment. A recently updated page may describe older work. For software, match the documentation to the actual version. For hardware, match the exact model and revision. Similar names do not establish compatibility.

Build a claim-and-source record

Use a short table with four columns: claim, supporting source, what the source actually says, and remaining limitation. The third column prevents a link from becoming decoration. Write a short paraphrase of the relevant evidence and record the section or page so you can return to it.

For example, a claim might be ‘This board has a camera connector.’ The source should identify the exact board and connector. That evidence does not also prove that your chosen camera cable fits, that the camera software supports your operating system, or that an object detector runs quickly enough. Those are separate claims requiring separate checks.

Keep observations distinct from interpretations. ‘The documentation lists a 3.3 V GPIO interface’ is a documented observation. ‘Therefore this 5 V sensor can be connected directly’ is an inference, and in this case a dangerous one. A correct premise can still lead to a wrong conclusion if an electrical compatibility step is skipped.

A research prompt with explicit evidence rules

Weak prompt: ‘Tell me the best computer for an AI robot and add sources.’ It invites an unsupported winner and may attach citations after the answer has already been chosen.

Improved prompt: ‘Help me investigate whether local camera processing suits a small indoor educational robot. First identify five decision questions. For each, name the type of primary source needed. If you can browse, open sources and attach each factual claim to the page that supports it; if you cannot browse, say so and provide search terms only. Separate documented facts, assumptions, and your interpretation. Do not choose hardware until the requirements are clear.’

The instruction defines what browsing must mean: access to actual pages, not an assertion that the model remembers a source. If your tool lacks browsing, you can open documents yourself and supply relevant excerpts with their titles and dates. The assistant can help analyze those excerpts, but it has not verified the rest of the document.

Follow-up prompt: ‘This citation lists the connector but says nothing about model speed. Remove the speed conclusion and identify a benchmark that would actually test our camera resolution and model.’ Verification: Open every cited page, locate the supporting passage, and check whether the scope matches the sentence. A source existing is only the first check.

Handle disagreement and write cautiously

Sources may disagree because they use different versions, workloads, definitions, or measurement methods. Do not ask the assistant to average incompatible numbers. First compare what was measured. Frames per second at one image size and model setting may not be comparable to a result using another pipeline.

When evidence is incomplete, state the missing condition. ‘The connector is documented, but this camera-and-software combination still needs confirmation’ is more useful than a confident yes. A conclusion can be conditional: choose local processing if the tested latency and memory use meet the project’s limits; otherwise narrow the model or move processing elsewhere.

Ask for counterevidence: what source or test would weaken the current conclusion? This reduces the temptation to collect only supportive links. Keep a record of rejected claims and why they were rejected. For research using AI, UNESCO’s education guidance places human oversight at the center; in this workflow, that means retaining responsibility for evidence selection and interpretation.

Your final paragraph should answer the original question, cite its strongest support, and name the most important limitation. More links do not automatically make the answer stronger. A few relevant, inspected sources are more useful than a long bibliography no one has checked.

Important terms

Primary source
Direct evidence for a question, such as original research or official specifications.
Citation
A reference identifying the source for a claim.
Inference
A conclusion drawn from evidence rather than stated directly by it.
Scope
The conditions and boundaries within which a claim applies.
Counterevidence
Evidence that weakens or contradicts a proposed conclusion.

Mini project: Research one small technical claim

  1. Choose a narrow question about a board feature or a programming concept. State the exact model or version when relevant.
  2. Ask an assistant for subquestions and source types, then locate two appropriate sources yourself or through verified browsing.
  3. Record three claims with their supporting passages and limitations. Remove any claim the source does not actually support.
  4. Write a short answer distinguishing facts from interpretation. Finish when another reader can trace each important claim to evidence.

Common mistakes and debugging

  • Treating a generated source list as verified: open the sources and inspect their contents.
  • Using a current-looking page for an older version: match dates and exact versions.
  • Citing a broad homepage for a precise specification: link the supporting page or document.
  • Comparing unlike measurements: align workloads, definitions, and test conditions first.

Independent challenge

Find a claim for which two sources appear to disagree. Explain whether the conflict is factual or caused by different definitions, versions, or test conditions.

Check your understanding: 10 questions

  1. What makes a question researchable?

  2. Is a manufacturer’s source automatically neutral?

  3. What must you check beyond whether a citation exists?

  4. Why should facts and inferences be labeled separately?

  5. What should you do with incompatible benchmark numbers?

  6. In your own words, what does “Primary source” mean?

  7. In your own words, what does “Citation” mean?

  8. In your own words, what does “Inference” mean?

  9. In your own words, what does “Scope” mean?

  10. In your own words, what does “Counterevidence” mean?

Quiz answers

Reveal all 10 answers after your attempt
  1. It has a clear scope and can be answered using identifiable evidence or a defined test.
  2. No. It can provide authoritative specifications while still presenting promotional interpretations.
  3. Whether the source actually supports the attached claim under the relevant conditions.
  4. An inference adds reasoning that may be mistaken even when its supporting facts are correct.
  5. Compare their methods and conditions rather than averaging them or declaring a winner immediately.
  6. Direct evidence for a question, such as original research or official specifications.
  7. A reference identifying the source for a claim.
  8. A conclusion drawn from evidence rather than stated directly by it.
  9. The conditions and boundaries within which a claim applies.
  10. Evidence that weakens or contradicts a proposed conclusion.

Summary

Use AI to structure the investigation, then inspect primary evidence, track claim support, and keep conclusions within the sources’ limits.

Continue learning

Continue with USE06 to apply the same evidence-and-testing discipline when asking AI for code.

Choose a connected learning path

Sources and further reading

Prepared 2026-09-18. Draft; research workflow reviewed and cited guidance checked. No product recommendation or benchmark is claimed.

GO DEEPER

Extra reading & source documents

Optional reading alongside the lessons. These sources do not add to your course lesson count.

Explore the AI model guides →