See what an integrity desk would see — before you submit
Journals now screen submissions for the marks of assistant-written text, paper mills and fabricated citations, and they do it with black boxes. This screen shows you the same marks in the open: every flagged phrase, every rule, every source. It is a signal for you to act on, never a verdict on you — and your manuscript is read once and discarded.
What it looks for
Excess vocabulary — the words that rose sharply with LLM-assisted writing (the 2025 Science Advances study, data MIT), without the everyday ones: a rate per 1,000 words, the words shown.
Tortured phrases — "counterfeit consciousness" for artificial intelligence: the paper-mill fingerprints (Cabanac, Labbé & Magazinov 2021).
References — through ScholaRef: DOIs that belong to a different paper, and cited works Crossref holds a retraction or withdrawal notice for. A reference that could not be matched is shown, never counted: that is not evidence it is made up.
Pictures — the same image reused (also flipped, rotated, rescaled); files that declare themselves AI-generated (IPTC, content credentials, generator names).
Structure — sentence-length spread and connective openers, read with the vocabulary.
Data availability — the statement found, its dataset DOIs asked of doi.org; or a sentence to adapt.
Published work — passages of your text searched in OpenAlex and in the Smart Scholars registry, and the target journal's recent abstracts: the share of your five-word phrases found in published titles and abstracts, with each source and the matching words. Not full texts, and not Crossref's Similarity Check.
The level. Strong needs two independent families of signals; one family alone, however strong, reads Notable at most (from 9 Oct 2026).
What it is not
Not a detector that says "AI-written" with a percentage. Those exist, they are black boxes, and on technical English they misfire; a journal that uses one should still show you its evidence. This screen only shows what can be shown: the phrase, the rule it broke, the source of the rule. English-calibrated; for other languages the reference and picture checks still apply.
Privacy
The file is read on this server and deleted before the report is shown. The report keeps counts and the flagged phrases with a few words around them — never the manuscript. Its link is private to whoever holds it; nothing goes to any journal. The comparison with published work sends up to 6 passages of 12 words to OpenAlex, as searches, and nothing else of the text; the report keeps the matching words as the published source prints them. Authors' addresses and iDs, when given, go only to ROR and OpenAlex as lookups and are not kept. If you tick the second opinion, that text is sent to the provider you see named, and to nobody else.
Disclosing AI use
COPE and the ICMJE both say the same two things: an AI tool cannot be an author, and any use of one in writing, figures or analysis is disclosed in the manuscript. The form beside you makes the statement in plain words, the same words every Smart Scholars page uses. Disclosure is not a mark against a paper; an undisclosed leftover sentence is.
