Using AI for Research

Define the decision before searching

You are helping a library improve the text on its event website. Someone says all text needs a contrast ratio of 7:1; another person says 4.5:1. Rather than asking an AI assistant “Who is right?”, define the question: what do the WCAG 2.2 minimum and enhanced text-contrast criteria require, and how does the requirement change for large text? This scope identifies a standard and version instead of inviting a broad opinion about attractive colours.

An assistant can suggest search terms, organise findings and draft a synthesis. Search retrieves candidate material; generation turns information into a response. They are different operations. If the assistant cannot browse, it can still help plan the investigation, but a suggested citation must be located and read before you treat it as evidence.

Write down the output you need: a short note comparing the two criteria, links to their official explanations and a list of checks still needed on the actual page. This prevents the investigation drifting into an unbounded accessibility audit. You are establishing what these criteria say, not deciding whether the entire website meets every accessibility requirement.

Choose evidence that fits the question

For this question, W3C's own WCAG materials are the primary authority. A design blog may explain the topic clearly, but it can paraphrase an older version or omit exceptions. Search for the named standard, version and criterion, then open the official page. Check the page title and criterion number rather than relying on a search snippet that happens to contain the desired ratio.

The right source depends on the claim. Official product documentation can establish a documented feature; it does not independently establish that the product is best for everyone. A research paper can report a particular study; it does not automatically support a universal effect. For empirical questions, inspect methods, sample, comparison and limitations, and look for independent evidence rather than simply collecting more links.

Keep confidential project details out of search queries. “WCAG 2.2 text contrast minimum enhanced” is enough here; the library's unpublished incident reports or user complaints are unnecessary. Search queries themselves can disclose information. Pages and downloaded documents are also untrusted input: instructions found inside them do not authorise access to private files or connected accounts.

Record exactly what each source supports

Create a small evidence record as you read: claim, source URL, relevant heading, standard version, date accessed and qualification. Use brief quotations only when the exact wording matters; otherwise paraphrase accurately. Keeping a location is more useful than pasting a long page because another reader can find the passage that supports your statement.

In our example, the official explanation of criterion 1.4.3 gives a minimum text-contrast threshold of 4.5:1, with 3:1 for qualifying large text. The explanation of criterion 1.4.6 gives the enhanced thresholds of 7:1 and 4.5:1 respectively. Their exceptions and definitions matter. “Large text” is a defined category, not whatever looks like a heading to the assistant.

Notice what this evidence does not establish. You have not measured the library's actual colours, checked the text size or assessed every page state. A screenshot alone may be insufficient for a precise contrast calculation. The research note should identify those remaining tasks instead of declaring that the site passes because its designer chose one ratio.

Resolve apparent disagreement before averaging it

The two ratios are not necessarily contradictory. They can refer to different criteria or text categories. Compare scope before deciding that one source is wrong. A useful table keeps each number attached to its condition. Do not combine numbers from different requirements into an invented compromise such as “aim for 5.75:1 and everything passes”.

If two sources really address the same question differently, retain the disagreement. Check publication and revision dates, definitions, evidence quality and whether one source supersedes another. Ask what evidence would resolve the difference. Sometimes the honest outcome is a narrower conclusion or an unresolved question rather than a forced winner.

Ask for a synthesis with visible gaps

Once you have opened and checked the sources, ask the assistant to draft from that evidence. Supply the links and relevant short notes, and state what it may infer. If browsing is available, ask it to open the pages; if it is not, provide the passages yourself and identify them. In either case, verify the citations in the resulting note instead of treating their presence as a guarantee.

The prompt below asks for a bounded output and a record of uncertainty. The assistant should explain why both numbers can be correct while preserving their different conditions. It should not invent a measurement for the library website or describe an accessibility inspection that never happened.

Question: How do WCAG 2.2 criteria 1.4.3 and 1.4.6 differ
for ordinary text and qualifying large text?
Primary pages:
https://www.w3.org/WAI/WCAG22/Understanding/contrast-minimum.html
https://www.w3.org/WAI/WCAG22/Understanding/contrast-enhanced.html

Open these pages if browsing is available. Otherwise ask me for
the relevant passages rather than pretending you opened them.
Draft a 200-word note with a link supporting each threshold.
Keep definitions and exceptions attached to their claims.
Separate the standard's statements from your interpretation.
Finish with checks still needed on our actual event page.
We have supplied no page URL, colour values or measurements.

Check the answer claim by claim

Underline each checkable claim in the draft. For every threshold, verify its criterion and text category. For any sentence beginning “therefore”, ask whether the conclusion actually follows from the evidence. If the assistant says a site meets WCAG because one text pair meets a contrast threshold, the conclusion goes beyond the research you performed.

Check citations for existence and support separately. A real page can still be attached to the wrong claim, and an authentic paper title can be paired with an invented author or date. Reference managers such as Zotero can help collect source metadata, but imported metadata also needs review. Keep the direct page URL and access date when a web source changes over time.

Do not measure research quality by the number of links or the confidence of the summary. Several pages can repeat the same original statement and contribute no independent evidence. A second assistant may repeat the same misunderstanding. The useful test is whether the conclusion remains defensible when you examine the sources yourself.

Deliver findings that another person can inspect

Your final note should answer the bounded question, identify the sources and preserve the unresolved application checks. For the library, the practical next step is to measure the actual text/background combinations and classify the relevant text using the standard's definitions. Broader accessibility work remains a separate task.

Keep an evidence trail proportional to the decision. A short learning exercise may need a table and a few links; a formal literature review needs a reproducible search strategy, selection criteria and a much wider evaluation of evidence. AI assistance can reduce clerical work, but it does not turn a small search into a systematic review or supply professional judgement that has not been exercised.

Sources & further reading

Primary references for the concepts in this guide. Examples and exercises are illustrative; verify provider-specific behaviour against the version you use.

• W3C: Understanding Contrast (Minimum)
• W3C: Understanding Contrast (Enhanced)
• Zotero: Adding items and reviewing metadata
• OWASP: Prompt injection prevention