Articles on: How To Use Originality

Human-Written Content Flagged as AI: What to Do Next

A high AI score does not prove that content was generated by AI.


AI detection is probabilistic: it looks for patterns associated with AI-generated writing and estimates how likely those patterns are to appear in the scanned text. Human-written content can sometimes produce a high AI score, particularly when the text is short, highly structured, heavily edited, translated, or written in a predictable style.


Use the steps below to investigate the result before reaching a conclusion.


First, understand what the score means


An AI score is a confidence signal—not the percentage of words written by AI.
For example, a result of 90% AI does not necessarily mean that AI wrote 90% of the document. It means the detector found patterns that gave it high confidence that AI may have influenced the submitted text.
The score should be considered alongside:

  • The amount and type of text scanned
  • The detection model used
  • Editing and translation tools
  • Document history
  • The author’s previous work
  • Citations and source quality
  • Human review

No AI detector should be used as the sole basis for a disciplinary, employment, publishing, or academic decision.


Step 1: Scan the complete document


Short passages provide less context and are more vulnerable to uncertain results.
For the most reliable evaluation:

  • Scan the entire article, essay, or document
  • Use substantially more than the minimum word count when possible
  • Avoid scanning only an introduction or conclusion
  • Do not evaluate one isolated paragraph unless that is the only text available
  • Remove content that is not part of the author’s prose when appropriate

Introductions, conclusions, summaries, and other short passages are often highly structured. Scanning them alone can produce a different result from scanning the complete work.


Step 2: Check which model was used


Make sure the detection model is appropriate for the content.
Consider:

  • The language of the document
  • Whether the content is academic or technical
  • Whether it contains code, equations, or scientific terminology
  • Whether a multilingual or academic model is more appropriate
  • Whether two compared scans used different models

When comparing results, use exactly the same text and model. Otherwise, differences may be caused by the scan configuration rather than the content.


Step 3: Review editing and writing tools


Ask whether any software helped plan, translate, rewrite, paraphrase, or polish the text.
Examples include:

  • Grammarly
  • Microsoft Editor
  • QuillBot
  • DeepL
  • ChatGPT or another AI assistant
  • Generative rewrite features
  • Translation software
  • Automated summarizers
  • Tone or style rewriters
  • Predictive writing tools


These tools can significantly change sentence structure, vocabulary, rhythm, and predictability. A document may therefore contain AI-like patterns even when the author
wrote the original draft.


This does not automatically mean the author acted dishonestly. The important question is what assistance was used and whether that assistance was permitted in the relevant context.


A useful test

If an earlier draft exists, scan:

  1. The version before automated editing
  2. The version after automated editing
  3. The complete final document

Comparing these versions may show whether the score changed after a particular tool was used.


Step 4: Look for content that is naturally predictable


Some types of human writing are more formulaic than others.
Examples include:

  • Introductions and conclusions
  • Five-paragraph essays
  • Definitions and standard explanations
  • Recipes and instructions
  • Product descriptions
  • Legal or compliance language
  • Technical documentation
  • Scientific writing
  • Reference lists and bibliographies
  • Frequently repeated marketing phrases
  • Public-domain or historical text

These formats often use similar sentence structures and vocabulary across many documents. That predictability can sometimes resemble AI-generated writing.
Scan a longer and more representative portion of the author’s work before interpreting the result.


Step 5: Review citations and references separately


Bibliographies, quotations, citations, and standardized reference formats are highly structured.
When appropriate:

  1. Save the original scan.
  2. Create a copy of the text.
  3. Remove the bibliography, reference list, and long quotations.
  4. Scan the author’s prose separately.
  5. Compare the results.

Do not remove citations simply to obtain a lower score. This comparison is a diagnostic step intended to determine which part of the document influenced the result.


Step 6: Compare multiple writing samples


One document provides limited evidence.
When authorship matters, compare several substantial samples written by the same person. Ideally, use:

  • Three to five writing samples
  • Similar subject matter
  • Similar document lengths
  • Work created under comparable conditions
  • Samples known to predate widespread generative-AI use, when available

Look for the author’s normal patterns rather than expecting every document to produce the same score.
A sudden change can justify further review, but it is not proof by itself.


Step 7: Review document history


Document history can provide valuable context beyond an AI score.
Look for:

  • Gradual drafting over time
  • Revisions and corrections
  • Deleted and rewritten passages
  • Paste events
  • Multiple contributors
  • Large blocks inserted at once
  • Unusual automated typing patterns

If the work was created in Google Docs, the Originality.ai Chrome Extension may provide Writer Replay and typing-behavior insights.


Important limitations


Incomplete history does not prove AI use.
Writing history may be limited when:

  • The content was written in Microsoft Word and pasted into Google Docs
  • The document was imported or converted
  • Work began before revision history was available
  • The author drafted the text in another application
  • A contributor worked in a separate document
  • Portions were copied from the author’s own notes

Use document history as supporting evidence, not as a final verdict.


Step 8: Speak with the author


If the result is being reviewed by an educator, editor, employer, or client, give the author an opportunity to explain their process.
Useful questions include:

  • How did you plan and draft the work?
  • Where was the document originally written?
  • Did you use grammar, translation, or rewriting tools?
  • Do earlier drafts or notes exist?
  • Can you explain the argument or research in your own words?
  • Can you identify and discuss the sources?
  • Does the document’s revision history support the explanation?
  • Is similar earlier writing available for comparison?

A conversation about the work often provides more useful evidence than repeatedly scanning small passages.


Step 9: Rescan carefully


Before rescanning, confirm that you are making a valid comparison.
Use:

  • Exactly the same text
  • The same detection model
  • The same language settings
  • The same included or excluded citations
  • A complete, representative sample

Repeatedly rewriting sentences solely to lower the AI score is not a reliable investigation method. It changes the evidence and can make the writing less natural.


What not to do


Avoid these common mistakes:

  • Treating an AI score as definitive proof
  • Interpreting the score as the percentage written by AI
  • Testing only a short paragraph
  • Comparing different text across detectors
  • Ignoring grammar, translation, or rewriting tools
  • Assuming missing revision history proves misconduct
  • Making a major decision based on one scan
  • Asking someone to repeatedly rewrite authentic work until it “passes”
  • Relying on highlighted sentences without reviewing the entire document



Use a consistent process for every author:

  1. Preserve the original document and scan.
  2. Confirm the text and model that were used.
  3. Review the complete work.
  4. Identify permitted and prohibited writing tools.
  5. Examine drafts, sources, and revision history.
  6. Compare prior writing samples.
  7. Discuss the work with the author.
  8. Document the evidence considered.
  9. Apply the same standard to comparable cases.
  10. Allow an appropriate review or appeal process.

Your policy should explain whether tools such as grammar correction, translation, predictive text, or generative rewriting are permitted. Many disputes arise because “AI use” has not been clearly defined.


When to contact Originality.ai support


Contact support if:

  • The same text produces an unexpected error
  • You are unsure which model was used
  • A scan appears incomplete
  • Highlighted results do not load
  • You need help sharing the scan
  • Results change unexpectedly without changes to the text or model
  • You believe there may be a technical issue


Include:

  • A shareable scan link
  • The detection model
  • Text language
  • Approximate word count
  • A description of any editing or translation tools used
  • Screenshots of the result
  • The browser or integration used for the scan

Do not send confidential content unless you are authorized to share it.


Quick decision tree


Was the complete document scanned?

  • No → Scan the complete document.
  • Yes → Continue.


Is the text short, formulaic, technical, historical, or mostly citations?

  • Yes → Scan a longer, more representative sample and review citations separately.
  • No → Continue.


Were grammar, translation, paraphrasing, or rewriting tools used?

  • Yes → Compare versions from before and after those tools were applied.
  • No or unknown → Continue.


Was the correct model used for the language and content type?

  • No or unknown → Select the appropriate model and rescan the same text.
  • Yes → Continue.


Are drafts or document history available?

  • Yes → Review the writing process, revisions, paste events, and contributors.
  • No → Use other evidence; absence of history is not proof of AI use.


Are previous writing samples available?

  • Yes → Compare several substantial samples from the same author.
  • No → Continue with a contextual human review.


Does the result still appear inconsistent with the available evidence?

  • Yes → Preserve the scan and contact Originality.ai support.
  • No → Document the evidence and interpret the score as one signal among several.

Updated on: 24/07/2026

Was this article helpful?

Share your feedback

Cancel

Thank you!