PlagiatScanner

AI transparency · verification · KIT-14

Verify AI summaries against original sources

In brief: Split a generated summary into independently checkable claims. Map each claim to a precise passage in the original, compare meaning, strength of evidence, scope and omissions, and mark anything unconfirmed. Only then write your own account and cite the original. Generated prose is neither the evidence nor a substitute for reading the source.

Claim-by-claim verification worksheet

Fields for each proposition in a generated summary
Field Record
Generated sentence Unaltered candidate under review
Source location Page, section, table or paragraph
Status Supported, partly supported, contradicted or not found
Evidence type Observation, interpretation, hypothesis or recommendation
Scope Population, period, material and conditions
Omission Counterfinding, qualification or uncertainty
Author's account New wording supported by the original

Turn fluent paragraphs into atomic claims

A summary often combines result, cause and evaluation in one sentence. “The study proves that method A permanently improves performance” contains a measured effect, causation, duration and judgement about proof. The paper might report only a short-term association in a narrow sample. A matching phrase does not validate the whole sentence.

Start with verbs and quantifiers. Flag “demonstrates”, “causes”, “all”, “significant”, “long-term” and similar intensifiers. Search for the proposition, not just identical vocabulary. Tables, notes and limitations may change its meaning. Abstracts support orientation; detailed assertions require the relevant full-text location.

Preserve evidence strength and boundaries

Generated prose can smooth away caution. “May be associated under these conditions” becomes “leads to”. Compare modality, uncertainty and the status of the statement. Does it come from results or discussion? Is it a direct observation, author interpretation or a reference to another work? In the last case, the cited primary publication may need checking.

  • Record the studied group and selection procedure.
  • Separate measurement instrument from outcome.
  • Preserve timing and observation period.
  • Note control or comparison conditions.
  • Consider exclusions and sensitivity analyses.
  • Include limitations and rival explanations.

A concise account need not reproduce every detail, but it must not omit a condition that reverses or sharply narrows the meaning.

Classify the fault before rewriting

Not found: discard the claim; never invent a page. Partly supported: split the sentence and retain only the evidenced part. Overstated: restore uncertainty and scope. Misattributed: locate the primary source actually responsible. Decontextualised: reinstate the condition under which the finding holds.

Write your own account from notes with the generated paragraph closed, then check it again against the original. This reduces the chance of reproducing the same misleading structure with synonyms. Cite the work you read, not the language model.

Disclose the aid without implying certainty

A proportionate statement could say: “On 19 August 2026 I used [tool] for a preliminary orientation to article X. I checked every proposition against a page or section in the full text, removed unsupported or overstated claims, and cite only the original publication.” Add the affected chapter and location of the worksheet.

If the original was unavailable, mark the candidate unresolved and do not use it for detailed claims. Disclosure cannot replace source access or assessment-specific rules.

The AI transparency hub connects this record with prompt and model evidence. The literature verification workflow also checks bibliographic identity. For the writing boundary, consult avoiding plagiarism.

Worksheet audit

  1. Does every factual sentence have a specific source location?
  2. Were compound propositions assessed separately?
  3. Are uncertainty and scope preserved?
  4. Are contradictions and limitations represented?
  5. Was a secondary claim mistaken for a primary result?
  6. Was the author's version written independently?
  7. Does the citation identify the original?

Keep the worksheet with the source version used. Pagination may differ between a preprint and version of record. Record the version, DOI or stable address and access route so that the verification remains intelligible.

Do not outsource the final audit to the same generated summary. The key question is whether a reader can travel from your statement to the exact evidence and see the same degree of certainty. A transparent “not established by this source” is a valid outcome and should remain in research notes.

Check tables and figures as evidence

For numerical claims, inspect the table or figure, caption, unit, denominator, group and axis scale. A percentage may describe a subset and an error bar may not mean what the summary assumes. Record that reference in the worksheet. Compare result prose with captions; if they differ in emphasis, do not silently select the stronger account. Return to the method and calculation. An attractive chart never substitutes for checking what its denominator represents.

Sources

Check whether a summary has blended several publications. A result from work A and a qualification from work B can become a new proposition supported by neither. Map each clause to its own evidence. Where you synthesise, label it as your comparison and explain relevant methodological differences. Preserve conflicting findings rather than smoothing them into artificial consensus; “sources disagree” is a valid worksheet outcome and may be the most important conclusion.

Sources reviewed 4 September 2026.