PlagiatScanner

AI transparency · definition · KIT-16

Why AI output does not replace a scholarly source

In brief: A generated answer does not automatically provide a checkable origin, stable publication state or dependable passage for its assertions. Use it as a discovery or working aid, verify factual content in original publications and cite those works. Where the output itself is the object of research, it can become primary material that must be archived and described; it still cannot validate its own factual claims.

Source-status decision tree

  1. Must the statement establish something about the world, research or a work? Locate and cite a verifiable original; generated output is insufficient.
  2. Did the answer only support searching, structuring or wording? Record tool use as required and support factual claims separately.
  3. Are you studying this exact response? It is research material. Preserve prompt, system, date, displayed version and unchanged output.
  4. Does the system name literature? Treat each item as unconfirmed until record and content are checked.
  5. Can no original be found? Do not present the claim as a research finding.

What makes a source inspectable?

A source gives a reader an identifiable object whose context can be examined. A scholarly publication normally has authorship, title, publication venue, version and a stable identifier or catalogue record. Readers can evaluate method, data, reasoning and qualifications. Publication does not guarantee quality, but it makes the object available for criticism.

A chat response may vary when a prompt is repeated, blend sources or supply convincing detail without provenance. The path from training material to a sentence is generally unavailable to the user. Even displayed citations require independent checks. The answer may open a search route; it cannot shorten the evidence chain.

Three roles that should not be conflated

Discovery aid: the system offers terms or candidates. The result is a lead list, not a bibliography. Editing aid: it organises notes, translates or rewrites. Responsibility and source attribution remain with the author. Object of study: the response is analysed for errors, discourse or variation. The interaction then needs reproducible documentation and a defined sampling method.

Tool role and matching evidence
Role Evidence cited Process record
Discovery Verified original publication Search and verification log
Editing Original source of the proposition Use and revision record
Research object Archived concrete response Prompt, context, system, date and sample

When output becomes primary material

If a project investigates system responses, design data collection like other empirical material. Define prompt selection, number of runs, ordering, parameters, language, timing and analysis. One conversation cannot establish a general property of a model. Product changes and hidden system components limit reproducibility.

Archive responses unchanged and assign local identifiers. Do not include personal or confidential inputs. Distinguish observed output from your interpretation. Making the response primary material for this question does not make the statements inside it true.

From generated lead to citeable evidence

  1. Break the answer into checkable propositions.
  2. Extract key terms and possible originals.
  3. Search catalogues, subject databases, Crossref or official institutions.
  4. Match metadata and publication status.
  5. Read the original and preserve the relevant context.
  6. Write an independent account and cite the original.
  7. Disclose AI assistance separately where required.

An AI declaration is not a citation. Conversely, correct citation does not replace disclosure of material AI assistance. They answer different questions.

Use the literature verification workflow for practice. The AI transparency hub covers process evidence. The scholarly sources guide explains publication types and evidential purpose.

Frequent category errors

  • Citing a response although the underlying study is available.
  • Completing an invented title with plausible metadata.
  • Using a summary instead of reading the source.
  • Generalising about a system from one chat.
  • Archiving the prompt while leaving factual claims unsupported.
  • Treating the provider as author of scholarly findings.

The remedy has two tracks: an evidence chain for content and a use chain for process. Keeping them separate prevents both hidden assistance and fabricated authority.

If an institution requires a citation-like entry for an interaction, follow that local format, but do not mistake the entry for corroboration. It identifies research material or assistance; another source must still support external facts.

Identity does not equal truth

A DOI, ISBN or repository link does not guarantee that content is correct or suitable. It first answers “which object?”. Next ask whether publication status is current, what method was used, which claim actually appears and whether it fits your context. Keep three fields: identity confirmed, content checked and suitability assessed. All must be positive for the particular claim. Another passage in the same work may require a separate appraisal because source criticism attaches to use, not merely to documents.

Sources

Definitions deserve the same separation. Generated wording can blend several disciplinary traditions into one apparently settled formula. Locate definitions actually used in your field, state differences and justify the working definition adopted. Let the generated version guide discovery, not become the hidden origin of a central concept. Where a term remains contested, a documented range is more accurate than false uniformity.

Sources reviewed 4 September 2026.

Explore further