A signal is not a verdict
AI detector limitations: what false positives and false negatives mean
AI-text detection is an estimate based on patterns, reference data, and the text available to the analysis. A responsible result includes uncertainty and context instead of pretending that a score is proof.
False positives are possible
A false positive happens when human writing contains patterns associated with machine-assisted writing. Formal academic prose, standardized business language, language-learning writing, translation, and extensive editing can all look unusually regular to an automated detector.
A high signal should therefore start a conversation or a closer review. It should not, by itself, trigger a disciplinary, employment, legal, or academic conclusion.
False negatives are possible too
A false negative happens when AI-assisted writing is not flagged. Rewriting, paraphrasing, translation, short passages, mixed human and machine text, and writing outside the available evidence can reduce the signal.
No detector can guarantee that every AI-assisted passage will be found. A low signal is not proof that a person wrote every word without assistance.
Inputs that need extra caution
- Very short text, where a few sentences provide little evidence.
- Translated, multilingual, or heavily edited text.
- Code, formulas, lists, citations, tables, and template-driven writing.
- Specialized domains or styles that differ from the reference evidence.
- Scanned documents or files whose text extraction is incomplete.
Always check the product’s coverage and reliability notes. A number without sufficient analyzable text is not a strong whole-document conclusion.
Do not attribute text to a named product
Overlapping writing patterns cannot reliably identify ChatGPT, Claude, Gemini, DeepSeek, or another specific vendor. Even when a result shows a stronger AI-associated signal, it does not reveal the prompt, tool, author, editing history, or source of the text.
“This passage contains signals worth reviewing” is a review statement. “This proves a named person used a named system” is an attribution claim the text alone cannot support.
What to combine with a detector result
- The document’s drafts, revision history, notes, and source files.
- Citations, factual understanding, and the author’s ability to explain the work.
- Consistent policy, human review, and a way to appeal consequential decisions.
For high-stakes decisions, an automated detector should never be the sole evidence.