methodsguide

How to Verify Sources in AI-Assisted Research

A claim-led verification procedure designed for volatile product and research claims.

Three annotated source cards and a magnifying glass show a research verification workflow.
Last reviewed 2026-10-16

An AI research answer can cite a real page and still widen what that page says. To verify it, isolate one consequential claim, open the cited source, find the supporting passage, and compare the scope of each. In a bounded test, an assistant summarized OpenAI’s prompt-caching guidance accurately in most respects but made a discount sound broader than the source supports.

What we found: the answer said prompt caching can reduce “input-token costs by up to 95%.” The source ties that discount to reused cached input tokens, not necessarily every input token in a request. We classified the claim as PARTIALLY_SUPPORTED and narrowed the wording. This one synthetic test does not measure how often any model makes the same error.

One claim, opened and checked

On 15 October 2026, we gave gpt-5.6-sol a short synthetic developer question and a source packet summarizing OpenAI’s prompt-caching documentation. We asked for an answer of at most 100 words citing the page. The retained prompt and raw response show exactly what the model received and returned. We then opened the cited page independently on 16 October.

Check Observation
AI claim “Prompt caching can reduce input-token costs by up to 95%.”
Cited source OpenAI’s prompt-caching guide.
Source meaning Matching prompt prefixes can be reused; the discounted rate applies to cached input tokens.
Difference The AI omitted “cached/reused,” which could make the saving sound like a discount on all input tokens.
Classification PARTIALLY_SUPPORTED. The number and mechanism are real; the cost scope needs qualification.

A safer version is: “When a request reuses an eligible matching prompt prefix, the cached input tokens can receive a discount of up to 95%; other input tokens are billed normally.” The answer’s prefix-match and 1,024-token conditions were consistent with the source packet. The response also advised placing stable content first, a practical inference from prefix matching rather than a direct quotation. That sentence should be marked as advice, not source text.

A five-minute source check

  1. Copy the one sentence that would affect your decision. Split compound sentences into separate claims.
  2. Open the cited page, rather than trusting the link text or the assistant’s summary.
  3. Find the exact passage and note who published it, its date or version, and the conditions around the claim.
  4. Compare scope: numbers, populations, product versions, causal language, and exceptions.
  5. Mark SUPPORTED, PARTIALLY_SUPPORTED, UNSUPPORTED, INFERENCE, OUTDATED, or SOURCE_NOT_FOUND. Rewrite or remove the claim, then save the source and review date.

The free AI Source Verification Checklist gives you fields to repeat that check. For high-stakes decisions, have a qualified reviewer inspect the original evidence as well.

This article’s original contribution is the bounded claim check and a repeatable review procedure. It does not claim that one current product always performs better.

The working principle

Treat the model as a transformation layer, not a source of truth or an invisible author. The quality of the artifact depends on the decisions surrounding the prompt: which evidence enters, what the task excludes, how intermediate claims are inspected, and who owns the final judgment. A fluent answer is only a candidate artifact.

Our reusable control is an evidence packet with six fields: the decision to support, approved source material, known constraints, required output, evaluation criteria, and unresolved questions. This packet is deliberately more demanding than a clever sentence. It makes omissions visible and lets another person repeat the work.

Build the input before opening a tool

  1. Write the decision in one sentence. If the sentence contains “and,” consider splitting the work.
  2. Label every input as evidence, interpretation, constraint, or example. These labels stop an example from quietly becoming a fact.
  3. Record exclusions: claims the material cannot support, private information that must not be supplied, and conclusions already ruled out.
  4. Specify the artifact and its audience. “Summarize this” is not a professional deliverable; “produce a claim ledger for an editor deciding which themes merit follow-up” is.
  5. Define a small rubric before seeing output. We use traceability, specificity, contradiction handling, usability, and repair effort.

Key takeaway: Context is not volume. It is an explicit map of evidence, constraints, decisions, and quality criteria.

Work in inspectable passes

A single long response hides where the reasoning changed. Use bounded passes instead. First ask for an inventory without synthesis. Second request clusters with source identifiers beside every item. Third ask for counterevidence and missing information. Only then request the final form. Save the intermediate tables: they are often more useful during review than the polished answer.

At each pass, reject unsupported certainty. A useful response can say that the supplied evidence is insufficient. An answer that completes every field regardless of evidence creates more work because the editor must discover which sentences are inventions, assumptions, or harmless connective prose.

Review the artifact, not the fluency

Read once for the argument and again for provenance. During the provenance pass, underline every external or consequential claim and locate its support. During the judgment pass, ask whether the artifact makes a choice, whether alternatives were fairly considered, and whether its language could describe any competitor or project. Delete generic claims rather than decorating them.

Measure repair effort. Count unsupported claims, lost qualifications, collapsed disagreements, and sentences that require substantive rewriting. Speed to first draft is not the useful metric when the draft transfers verification costs downstream.

Failure modes and corrections

Evidence soup. A large upload is treated as context although the model cannot know which passage governs which claim. Correct it with source IDs and a claim ledger.

Premature synthesis. Themes appear before the material has been inventoried. Correct it by separating extraction, clustering, challenge, and writing.

Polished neutrality. The draft lists reasonable observations but makes no consequential choice. Correct it by naming the decision and forcing alternatives to compete against criteria.

Voice laundering. Generic phrasing is accepted because it sounds professional. Correct it with a banned-language list, concrete examples, and a final human rewrite.

Audit claims, not bibliographies

Extract every consequential claim into a ledger before checking links. Open the source and locate the exact passage; a plausible title is not support. Record whether the source supports the whole claim, a narrower claim, or only background. Check publication and update dates for volatile behavior. Identify who benefits commercially from the assertion. Search for material contradiction when the claim affects a recommendation. Then revise the sentence to match the evidence rather than shopping for a weaker citation. A source count is not a trust score: one current primary source can be better than ten derivative articles. Preserve access dates and the reviewer’s judgment so a future refresh knows what to reopen.

A compact worked example

Suppose six notes describe a fictional independent studio deciding whether to standardize its client discovery. Four praise consistent kickoff questions; one says the same questions suppress unexpected problems; one reports that teams cannot find the evidence behind the final recommendation. A direct summary might call “consistency” the theme. The staged method records two tensions instead: consistency versus discovery, and speed versus traceability. The decision becomes testable: standardize evidence capture and review criteria, but let interview questions branch when contradictory information appears. The resulting artifact cites N1–N6, records the dissent, and names the next test. No external fact or invented participant is required.

That example illustrates the editorial standard. The model may help inventory, cluster, or restate supplied material, but a person decides whether the tension matters and whether the proposed action follows. The source IDs make disagreement cheap to inspect. The limitations make the artifact safe to hand to a colleague without pretending the synthetic packet represents a market.

Definition of done for this workflow

Before release, another reader should be able to answer five questions without asking the author: What decision is being supported? Which evidence supports each consequential claim? What contradictory evidence remains? Which step involved editorial judgment? What would cause the recommendation to change? If any answer is hidden inside chat history, move it into the artifact. If a current product capability matters, record the visible version, account context, test date, and official source. If the work involves people, confirm consent, minimization, and redaction before sharing it with any service.

Limits

This method does not make model output factual, private, or original by default. Product behavior varies by model, interface, account, and date. Vendor documentation explains intended capabilities but is not independent evidence of quality. Do not upload confidential interviews or client documents without authority and an appropriate data-handling review. The method improves auditability; it cannot replace domain expertise.

Continue the work

Use the source-verification checklist when a claim leaves the supplied packet. The broader messy-research workflow shows how evidence becomes a decision rather than a prettier summary.

Editorial judgment

The durable advantage is not a longer prompt. It is a process in which evidence and judgment remain visible at every handoff. Use AI where transformation is reversible and reviewable. Keep consequential selection, verification, and authorship with the person accountable for the work.

Practical resources

Evidence ledger

Sources

  1. 01Prompt caching
  2. 02Prompt engineering overview
  3. 03Prompt engineering best practices

Links are descriptive and separated from editorial conclusions. Product behavior may change after the review date.

Written by

Raúl Mermans

Raúl Mermans creates and edits Token & Taste, documenting practical AI methods and reviewing the evidence behind its guidance.

About the author →

Next methods

Continue the work

Or browse Latest