6. Citation, Paraphrasing, Integrity, and AI

The decision

A citation is not decoration. It is a route by which a reader can inspect the evidence behind a claim.

If any link is missing, the claim is not yet auditable.

Learning outcomes

You will be able to:

• decide when and where a citation is needed;
• paraphrase meaning rather than rearrange wording;
• cite a source only for the claim it actually supports;
• use AI as an assistive tool without delegating verification or judgment.

1. Decide what needs a citation

Cite when you use:

• another author's idea, definition, model, or wording;
• a factual claim that a reader may reasonably verify;
• data, a result, an estimate, or a contested interpretation;
• a method or measure developed elsewhere;
• a claim about what “the literature” shows.

Usually do not cite:

• your clearly labelled research question or decision;
• your own observed result, when the paper has already identified it as such;
• genuinely common knowledge in the intended discipline.

When uncertain, ask: Could a sceptical reader reasonably ask, “How do you know?”

2. Place the citation at the claim

Ambiguous:

Peer feedback may improve writing, confidence, revision behaviour, and course participation. Several studies support these outcomes (Author, year).

Clearer:

Quantitative syntheses report improvement in writing performance, while qualitative studies also describe affective and social benefits (Author A, year; Author B, year). These outcomes should not be treated as the same construct.

The second version shows which claim the evidence supports and prevents one citation from carrying four unexamined claims.

3. Record the evidence boundary

Use a claim record:

Draft claim:
Source:
Exact location: page / table / figure / section
What the source directly shows:
My inference:
Boundary or rival explanation:

Worked record

Draft claim: AI-generated references require independent verification.
Source: Walters & Wilder (2023)
Location: abstract and results tables
Direct evidence: in 84 generated papers on 42 topics, 18% of GPT-4 references were fabricated;
                 24% of its real references contained substantive errors
Inference: fluent reference lists were not reliable in this tested setup
Boundary: historical model versions, prompts, and topics; not an estimate for every current system

The study supports a verification rule, not the timeless claim that “AI always invents sources” (Walters & Wilder, 2023).

4. Paraphrase the idea

Paraphrasing requires three changes:

1. representation — select the part relevant to your question;
2. structure — rebuild the logic for your paragraph;
3. language — write without looking at the original, then verify accuracy.

Weak paraphrase

Walters and Wilder found fabrication and substantive errors in citations generated by ChatGPT.

This changes little and does not explain why the finding matters.

Analytical paraphrase

In a model- and prompt-specific test, plausible formatting did not guarantee that a reference existed or that its metadata was correct; generated citations therefore need checking against the publisher record or a trusted index.

The citation is still required because the evidence is borrowed, even though the wording and analytical use are new.

Use quotation only when wording is evidence

Quote when the exact language is itself being analysed, unusually precise, or historically important. Otherwise, paraphrase and explain. Always preserve page or location information while taking notes.

5. Synthesise without citation dumping

Weak:

Peer feedback improves writing (A, 2019; B, 2023; C, 2024).

Stronger:

Review-level evidence is broadly positive, but estimates and reported mechanisms vary with the comparison group, task structure, and outcome measure (A, 2019; B, 2023; C, 2024).

The stronger sentence tells the reader why the sources appear together.

6. Use AI by risk, not convenience

First follow your institution, programme, supervisor, publisher, and data-protection rules. Where use is permitted:

The central risk is not only false text. AI can make researchers feel that a topic has been understood while hiding alternative questions or viewpoints—the “illusion of understanding” described by Messeri and Crockett (2024).

7. Apply the AI verification protocol

For every AI-assisted factual statement or reference:

• locate the original source independently;
• match author, title, year, venue, and DOI/URL;
• open the source rather than trusting a generated summary;
• find the exact passage, table, figure, code, or data;
• check whether design and population support your wording;
• compare at least one independent source for important claims;
• record the tool, date, task, and material retained where disclosure is required.

Do not ask the same system that invented a reference to certify that the reference is real.

8. Disclose use precisely

Adapt this to local policy:

I used [tool/version] on [date] to [bounded task].
I did not use it as an evidentiary source.
I independently verified [facts/citations/analysis] against [records].
I remain responsible for the argument, interpretation, and final text.

Avoid vague statements such as “AI was used to improve the paper.” Name the actual task and control.

Practice: trace one paragraph

1. Choose the paragraph with your strongest claim.
2. Underline every borrowed fact, idea, measure, and inference.
3. Connect each to a claim record and exact source location.
4. Rewrite one paraphrase from notes without viewing the original wording.
5. Add the strongest boundary the source requires.
6. If AI assisted, complete the verification and disclosure record.

Quality check

• every important claim has an inspectable source or is clearly your result;
• citations sit beside the clauses they support;
• source notes distinguish quotation, summary, and your analysis;
• paraphrases preserve meaning without copying structure;
• AI output is never treated as evidence;
• confidential or restricted material has stayed in approved systems.

Next: revise the research decisions before polishing the sentences.