Editorial guide · Pre-publication review
AI detector guide for writers and editors
Define what AI use and disclosure the assignment permits, then use a detector to locate language signals worth reviewing. Editorial judgment must still examine facts, citations, original sources, version history, and the writer’s explanation instead of handing publication responsibility to a heuristic score.
01 · Define review standards
Personal writing, branded content, journalism, research, and commissioned work carry different responsibilities. Editorial policy should separate assistance, generation, fact-checking, bylines, and required disclosure.
State whether AI may support ideation, transcription, translation, grammar revision, summarization, images, or passage generation. Raise verification requirements for journalism, health, finance, research, or brand claims. A detector cannot replace a contract, editorial standard, or platform policy.
Record the tools, stages, prompts or input sources, adopted material, and human revisions. A specific account is more reviewable than a general statement that “AI was used,” and helps distinguish assistance, collaborative editing, and substantial generation.
Keep the assignment brief, interview or research notes, source list, first draft, version history, editorial comments, and final file. A continuous record explains how the work developed and reconstructs review when detection, attribution, citation, or factual accuracy is challenged.
02 · Run a writing screen
GenDetect is a free local-browser heuristic AI writing detector demo. It reads pasted plain text, not a PDF or document file, and does not output a calibrated probability of authorship.
Paste complete plain text when possible instead of selecting a few suspicious sentences. Longer context provides more vocabulary-diversity, repeated-phrase, sentence-length, punctuation, and character samples. Tables, notes, and layout still need review in the source document.
Open the free AI writing detectorReview repetition, sentence length, punctuation, and character-structure factors, then compare the genre, house style, template, translation, and writer’s previous work. Fixed formats, second-language writing, or short samples can raise a score, while a low score cannot prove fully human authorship.
The current tool cannot identify ChatGPT, Claude, or another specific model and cannot determine a writer’s identity. Asking for a humanizer, deliberate errors, or repeated synonym replacement damages the text and does not establish reliable provenance.
03 · Perform editorial verification
Fluent copy may still contain invented facts, broken citations, concealed copying, or unsuitable sources. A detector score covers language patterns, not fact-checking, plagiarism, or provenance.
Check that names, dates, numbers, quotations, papers, links, and captions exist and support the text. Seek independent primary sources for high-risk claims. Return untraceable citations, implausibly precise unsourced data, and nonexistent references to the writer for clarification.
Review the outline, interview notes, drafts, and revision history for coherent development, understandable changes in voice, and a writer who can explain important choices. A style shift is a question to investigate; it cannot by itself prove AI generation.
AI detection examines patterns inside the text, while plagiarism checking searches for matching or similar sources. They answer different questions. When needed, also verify original files, content credentials, reverse-image sources, and publication dates instead of treating a low AI score as an originality check.
Compare AI detection and plagiarism checking04 · Make the publication decision
Publishing or rejecting work affects writers, readers, and institutional trust. Editors should record multiple sources of evidence, hear the writer’s response, and choose disclosure and correction practices for the content risk.
Review detector output alongside the brief, writer explanation, process records, fact-checking, and source evidence. Do not set an automatic rejection threshold. False positives and false negatives both occur, so a consequential decision needs a responsible editor who can explain the reasoning.
Use the publication policy to state which stage involved AI, what a person verified, and whether images, translation, or summaries require separate labels. Apply the same rules to staff, freelancers, and submissions, with a correction, appeal, or clarification path.
Before giving unpublished copy, interviews, or personal data to any detector, confirm its data flow, access scope, and retention period. GenDetect currently processes input locally in the browser; a future backend would require updated privacy and methodology disclosure first.
Read the current privacy notice