New: Boardroom MCP Engine!

Ready to put this into action?

Get the complete AI Integration PlaybookPractical AI implementation guide — prompt engineering, workflow automation, and ROI frameworks.

Article 047 · Part 5

Research Papers, Sources, and Citations Without Made-Up References

Find the source, inspect the relevant evidence, and connect it to the exact claim you want to make.

By Randy Salars · Published

On this page
  1. Narrow the question before collecting papers
  2. Search real discovery systems
  3. Inspect before adding a source to the argument
  4. Examine a five-source working bibliography
  5. Build a claim-to-source matrix
  6. Write a synthesis rather than five summaries
  7. Verify the reference list as a separate task
  8. Practice: build your own five-source record

Find the source, inspect the relevant evidence, and connect it to the exact claim you want to make.

A bibliography can look convincing before you open a single reference. It has author names, dates, journal titles, and identifiers. Then one title leads nowhere. Another paper exists but says something different. A third was never read beyond its abstract.

Research quality depends on more than the appearance of references. A citation has to identify a real source and support the statement beside it. AI can help refine a question, suggest search terms, or organize your notes. Those tasks become useful when you verify the underlying material.

Think of a reference as a route back to evidence. If another reader cannot follow that route, or arrives at evidence that does not support your sentence, the citation needs repair.

Narrow the question before collecting papers

“AI and education” is too broad for a short research paper. Narrow it by population, activity, outcome, and scope.

For this article’s worked example, use:

What evidence can guide the design of AI-assisted study sessions that preserve independent recall, and what remains uncertain about their effectiveness?

This question has two related parts. Research on learning can inform the study activity. Research on AI tutoring can inform claims about a particular implementation. A study without AI cannot directly establish that a chatbot improves learning, even if its findings help you design a better practice session.

Decide what kinds of sources belong. An empirical study reports a research investigation. A practice guide interprets a body of evidence for use. A competency framework sets educational goals. All can be relevant, but they contribute different forms of support.

Search real discovery systems

Ask the assistant for terms and variants:

Help refine this research question and produce search terms for a library catalog or scholarly index. Separate retrieval practice, delayed retention, AI tutoring, and independent performance. Treat any suggested reference as an unverified lead until I locate it. Do not generate a bibliography from memory.

Search through your institution’s library, appropriate scholarly indexes, and original publishers or repositories. Use exact titles to resolve a lead and identifiers to check the record. For example, the 2006 retrieval-practice paper below has a record in PubMed that identifies its authors, journal, pages, and DOI. PubMed record for Roediger and Karpicke

A discovery record helps establish identity. It does not replace reading the relevant methods and results. If only an abstract is accessible, record that limit and do not claim to have inspected details it does not contain. Ask a librarian about lawful access when needed.

Inspect before adding a source to the argument

Record who created the source, what it investigated or proposed, and where the relevant evidence appears. For a study, note the participants, task, comparison, outcome, and timing. Check the publisher record for corrections or other status notices when preparing the final paper.

Do not confuse an article’s publication date with a website’s update date. Do not assume a manuscript, conference abstract, and final journal article are three separate studies. Identify versions and use the one appropriate to the claim.

Read enough surrounding context to preserve qualifications. Searching within a paper for one favorable phrase can miss the condition that limits it. Keep direct quotations brief, exact, and attached to their location.

Examine a five-source working bibliography

The following is a checked, selected reading set for this example, not a systematic review or a claim to cover all current research. Entries use compact editorial formatting; convert them to the style required by your assignment. The annotations state the role and limits of each source.

1. Roediger, H. L., III, and Karpicke, J. D. (2006). “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention.” Psychological Science, 17(3), 249–255. DOI: 10.1111/j.1467-9280.2006.01693.x. Undergraduate experiments compared studying prose passages with retrieval tests. The delayed outcomes support studying retrieval practice as a learning activity. This is not an AI study. Inspect Experiment 1 and Experiment 2, including their different test delays. Full paper

2. Karpicke, J. D., and Roediger, H. L., III. (2008). “The Critical Importance of Retrieval for Learning.” Science, 319(5865), 966–968. DOI: 10.1126/science.1152408. College students learned Swahili–English word pairs under different continued-study and continued-testing conditions. The one-week outcome informs questions about stopping practice after an initial correct response. It does not establish gains in every kind of subject performance. Inspect the condition descriptions and delayed-test results. Full paper hosted by MIT

3. Kestin, G., Miller, K., Klales, A., Milbourne, T., and 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, article 17458. DOI: 10.1038/s41598-025-97652-6. A randomized crossover study involving 194 Harvard undergraduate physics students evaluated a custom tutor across two lessons. Reported immediate learning gains favored that tutor condition. The setting, design, topics, and outcome timing limit claims about generic chatbots or lasting mastery. Inspect Results and Discussion. Full article

4. Pashler, H., and colleagues (2007). “Organizing Instruction and Study to Improve Student Learning.” Institute of Education Sciences, NCER 2007-2004. This practice guide summarizes recommendations with evidence ratings, including spacing, worked examples, and quizzing. It helps connect research to instructional design. It is guidance based on the evidence available at publication, not a new experiment with an AI tool. Inspect the recommendations and their supporting discussions. Official full guide

5. UNESCO (2024; official overview updated January 16, 2026). “AI Competency Framework for Students,” publication overview. The overview describes competencies and progression levels, including human-centered judgment. This reading set uses the overview for educational context, not as evidence of a measured learning effect. A detailed claim about the full framework would require consulting the linked publication itself. Official overview

Notice the last entry’s scope. It identifies the page actually used. Citing a full publication after reading only its overview would obscure what you inspected.

Build a claim-to-source matrix

Before drafting, connect each intended claim to evidence:

Intended claimSuitable support in this setLimit to include
Retrieval practice can support later recallSources 1 and 2State the tasks and populations studied
A designed AI tutor can be evaluated experimentallySource 3One implementation and study context
Study sessions can draw on evidence-informed design guidanceSource 4Guidance is not a direct product evaluation
Student AI education includes judgment and responsible useSource 5Educational goals, not an outcome trial
Every chatbot improves long-term achievementNoneRemove or replace the claim

The final row is a valuable research result. It identifies a sentence you cannot support with this set. Do not attach the nearest-looking citation just to complete the paragraph.

Write a synthesis rather than five summaries

A synthesis explains how sources relate. Here, the learning studies motivate including retrieval in practice. The AI study shows one way a designed tutoring system was evaluated. The guide and framework provide instructional and educational context.

A defensible inference is that a study workflow can combine source-based explanation, learner attempts, feedback, and an independent check. This is a design proposal informed by the selected sources. It is not a tested conclusion that this exact workflow produces a particular gain.

A useful gap is the difference between immediate success with a tool and delayed performance without it. You could make that distinction central to your paper rather than hiding it beneath broad enthusiasm or broad dismissal.

When sources disagree, compare their question, population, implementation, and outcome before declaring a contradiction. Different findings may concern different tasks.

Verify the reference list as a separate task

Check names, title, year, publication venue, pages or article number, and identifier against the source record. Reference software can help organize those fields, but imported metadata still needs review.

Apply the required style after establishing accuracy. Formatting cannot repair a nonexistent paper or an unsupported claim. Keep your notes and quotations attached to the correct source when moving paragraphs between drafts.

For a more advanced project, record search dates, databases, exact queries, and inclusion and exclusion rules. That makes the scope of the search reviewable. Do not call a selected reading list a systematic review without the method that such a claim requires.

Practice: build your own five-source record

Use this reading set or a teacher-approved question. Create an annotated bibliography and claim matrix. Mark what you read for each source: full paper, relevant section, or overview. Identify one tempting claim that the set cannot support and revise it.

Completion check: Every listed item is real, relevant, and represented within what you inspected. Your claims have source locations, your references have checked metadata, and your synthesis distinguishes findings from guidance and your own inference. No unread or invented reference is presented as verified evidence.

Get the AI Dispatch

Weekly insights on ai & technology — delivered to your inbox. No spam, unsubscribe any time.

Want to choose specific topics? Customize your interests