Teachers, editors and managers increasingly ask the same question: did a human write this? AI detectors promise an answer, and routinely deliver confident wrong ones.
Updated August 15, 2026, after a fresh review, with updated visuals and links.
Here's how to actually evaluate suspicious text: what detectors measure, when to trust them, and the human signals that work better than any scanner.
Short version: Understand what detectors measure first, then run the text through two detectors, not one. Details below.
Step by step
- Understand what detectors measure. Detectors score 'perplexity' and 'burstiness': how predictable the word choices are and how much sentence lengths vary. Human writing jumps around; AI writing glides. That's the entire trick, and it's statistical, not forensic.
- Run the text through two detectors, not one. Free tiers of GPTZero, Copyleaks or similar give a probability score. Two tools agreeing on 'likely AI' means more than one confident verdict. Never treat any single percentage as proof.
- Know the false positive problem. Formal, careful human writing trips detectors constantly: legal summaries, academic abstracts, and non-native English all score 'AI-like'. Studies have shown real student essays flagged above 90 percent. This is why schools that punish based on detectors alone keep losing appeals.
- Check the content, not the style. The stronger test: verify claims. AI writing often contains confident vagueness, citations that don't exist, and generic examples that fit any context. Ask the author to explain a specific paragraph aloud; genuine writers defend their work instantly, copiers stall.
- Look for process evidence. Google Docs and Word track version history: a document written over three days of edits looks different from one pasted in a single block. When stakes are real (grades, jobs, publication), ask for drafts or history rather than trusting a scanner.
- Have the honest conversation. If context allows, just ask. 'I use AI to outline and then write myself' is an increasingly normal and defensible workflow, and different policies apply. The detector arms race matters less than agreed rules about acceptable use.
Small tweaks, big difference
- Watermarks exist in theory: some AI providers embed statistical patterns in output, but no consumer tool reads them reliably yet.
- Rewritten AI text ('humanized' through a paraphraser) defeats detectors easily, which is the final reason to treat scores as hints.
- Your own familiarity is a detector: sudden style changes from a known writer say more than any website.
FAQ
Is there a reliable AI detector?
No detector is reliable enough for high-stakes decisions alone: published error rates are significant in both directions, and every provider's own documentation admits it. Use scores as one signal among several.
Can I check images for AI the same way?
Image detection is poorer still; look for metadata, reverse image search results, and visual artifacts (hands, text, reflections) instead of trusting image detectors.
Keep reading
- How to Create Videos With AI (Sora, Veo, and Friends): From Text Prompt to Finished Clip
- Quadratic Equation Solver: Roots for ax² + bx + c = 0 (Free Tool)
- How to Customize the Windows 11 Taskbar and Start Menu (Make It Yours in 10 Minutes)
Treat detectors as smoke alarms: worth hearing, never the whole fire investigation. Evidence of process beats probability scores every time.
