How to Detect AI-Written Text

Before reaching for a tool, scan the text for these patterns:

• "It's worth noting…" - And other hedge phrases ("importantly," "notably," "critically").
• Three-pronged structure - Every paragraph has exactly 3 bullet points or 3 supporting sentences.
• Balanced both-sides framing - No opinion, no stance, both perspectives presented neutrally.
• "Conclusion: X is multifaceted" - Vague summaries that say nothing.
• Uniform sentence length - 18-25 words each, very few short snappy sentences.
• No typos, no asides, no parenthetical jokes - Too clean.
• "It's important to remember…" - Meta-commentary about how to think about the topic.

The deeper tell is absence: no specific anecdotes, no named people or places that check out, no claims precise enough to be wrong. Human writers commit to details; models trained to avoid hallucination learned to stay vague. A 1000-word essay with zero verifiable specifics deserves suspicion regardless of what any scanner says.

The major AI detectors in 2026:

• Turnitin - Institutional standard. Used by universities. Cited at ~70-85% accuracy on unmodified GPT-4 output.
• GPTZero - Consumer-facing. Decent on chat-model text, weaker on humanized text.
• OriginalityAI - Popular for content marketing. Best at detecting humanized AI text.

For a quick free check, our own AI detector gives you a probability score with the usual caveats attached. None are perfect. False positive rates are 3-10%, and they're not evenly distributed - Formal academic prose, non-native English, and heavily edited text all flag more often. Don't accuse anyone based on a detector alone.

If three detectors all say "likely AI," you can be more confident. If one says yes and two say no, treat it as inconclusive. Disagreement is meaningful.

Also test in segments. Paste the document in 300-word chunks rather than whole: mixed documents (human writing with AI-generated sections spliced in) are increasingly common, and a whole-document score averages the signal away. A document that scores 40% overall but 95% on paragraphs three and four is telling you exactly where to look.

The strongest detection evidence usually isn't in the text - It's around it. Before judging prose style, check:

• Document version history (Google Docs, Word online): human essays grow over hours with messy edits; pasted AI text appears in one or two large blocks.
• Consistency with known writing: compare against the person's previous emails or essays. A sudden jump in vocabulary and polish is more diagnostic than any scanner.
• The citations: look up two or three. Hallucinated or subtly wrong references are the most common hard evidence of careless AI use.

This step costs ten minutes and produces evidence you can actually act on, which detector percentages are not.

If detection matters (academic, hiring), ask the writer for:

• Earlier drafts (with version-controlled timestamps if possible).
• Notes, outlines, scratch work.
• A few paragraphs written live or in front of you on a related topic.

People who actually wrote the text can produce these. Pure AI-paste-jobs cannot. The live-writing sample is the strongest of the three: ask a candidate or student to spend ten minutes extending their own argument in front of you. Someone who wrote the original picks up mid-thought; someone who didn't produces text in a visibly different voice, and the comparison settles the question better than any percentage.

Things detectors can't reliably catch in 2026:

• AI text written iteratively (multiple prompts, edits) by a human who treats the AI as a co-author.
• AI text run through good humanizer tools (catches dropped to ~50% in our 2026 testing).
• Short text (under 200 words). Detection needs statistical signal that's hard to extract from a paragraph.
• Non-English text (most detectors are English-focused).

If detection is critical, structure assignments around in-class writing or live oral defence.

An editor receives a guest post that feels off. Manual scan: every paragraph runs three sentences, two "it's worth noting"s, zero concrete examples - Suspicion rises. Detector pass: GPTZero says 88% AI, OriginalityAI says 91%, a third tool says 45%. Two strong signals, one dissent.

Chunked testing shows the intro and conclusion score human while the body scores heavily AI - A classic sandwich job. Citation check: one of three linked studies doesn't exist. The editor replies asking for the author's outline and a short revision handled live over a call. The author goes silent. Case closed without a single accusation being made - The process did the work.

Note what carried the decision: converging evidence plus a fabricated citation, not any single score. That's the standard to hold yourself to before acting on a detection.