AI detectors say they’re 99% accurate. Real-world tests say otherwise. Here’s the honest playbook for telling AI-generated writing from human work, without burning a contractor or trusting a tool that’s lying to you.
AI detectors claim near-perfect accuracy, but in real-world testing they routinely flag human writing as AI and miss AI text written with the right prompts. The reliable approach is a mix of human judgment on seven specific patterns, smart conversation with the writer, and (where it really matters) provenance tools that track where content came from.
Every AI detector on the market right now claims something between 95% and 99.98% accuracy. [1][2] If those numbers were true, none of us would have a problem. We could just paste the suspicious essay into Pangram or GPTZero, get a verdict, and move on with our lives.
The numbers are not true in the way you’d reasonably interpret them. They’re true on test sets the detector vendors built. In the wild, they’re considerably less reliable, and false positives (human writing flagged as AI) are common enough that universities, recruiters, and editors have been burned by them. [3]
Stanford researchers showed in 2023 that the major detectors were biased against non-native English writers, flagging their work as AI-generated more than half the time when it wasn’t. The detectors have improved since then. The bias hasn’t been eliminated, just narrowed. If you’re making a decision that affects someone’s grade, job, or contract on the basis of a detector score, you’re operating with worse evidence than you think you are.
Detectors look at a couple of things: perplexity (how predictable the next word is) and burstiness (how much sentence length and structure varies). [1] AI tends to be smoother and more uniform than humans on both. That’s the signal they’re chasing.
Three reasons it’s a losing fight:
Reason one: models keep getting better at sounding human. GPT-5.5 and Claude Opus 4.6 produce text that looks almost identical to a sharp human writer’s first draft. The training data for the detectors is two model generations behind by the time you’re using them.
Reason two: a tiny prompt change defeats most detectors. Ask Claude to “vary your sentence length, use contractions, and include one mild opinion in each section” and the detector signal drops dramatically. Anyone who’s been using AI for six months knows this trick.
Reason three: some humans naturally write in a uniform, formal register. Engineers documenting code. Lawyers drafting contract terms. Non-native English speakers using formal English to be safe. These writers get flagged constantly because their actual writing pattern matches what the detector was trained to call “AI.”
Detectors are bad. Human pattern recognition, trained on the right things, is much better. After reviewing thousands of pieces of submitted writing for our courses, here are the seven patterns I check first.
The long horizontal dash, technically called the em dash, is GPT-4 and GPT-5’s signature punctuation. ChatGPT especially loves it. If a piece of writing has more than two em dashes per page and the writer is not Cormac McCarthy, that’s a strong signal. The fix on the AI side is trivial (one line in the prompt) but most people who paste output unchanged don’t bother.
These are AI hedge phrases. Real writers either say the thing or don’t. AI introduces it with a softening preamble. If you see two of these in three paragraphs, you’re reading AI.
Look at any bulleted list. Real writers vary the structure. Some bullets start with a verb, some with a noun, some are full sentences, some are fragments. AI lists are often perfectly parallel: every bullet a noun phrase, every bullet exactly the same grammatical shape. Too tidy.
Real human writers rarely end a piece by summarising what they just wrote. They end with a forward-looking thought or a punchy line. AI almost always wraps with a recap. If the closing paragraph starts with “in conclusion,” “to summarise,” or “as we’ve explored,” your detector should be in your eyebrow.
“AI tools can help you save time on email” is the AI version. “I use this exact prompt every Tuesday before my team meeting and it saves me 20 minutes” is the human version. AI defaults to abstraction unless you push it. Ask yourself: does the writing name specific tools, specific people, specific numbers, or does it stay one level above the actual world?
Open the document. Squint. Are all the paragraphs the same height? That’s an AI tell. Real writing has rhythm: a one-line paragraph, a four-line paragraph, a seven-line paragraph. Uniform paragraphs are a structural fingerprint.
“While there are advantages and disadvantages to consider, the optimal approach may vary depending on your specific context.” That sentence says nothing and reads fine. AI defaults to balance. Real writers take positions, even if they qualify them. If you can read a 2,000-word article and not know what the writer actually thinks, that’s a signal.
Here’s a single technique that beats most detectors when you genuinely need to know whether a person wrote something. Ask the writer a follow-up question.
Specifically: pick one specific claim in the piece (a number, a date, a tool feature, a quoted statistic) and ask them where it came from. Ask them what changed since the date they cited. Ask them to describe the tool’s interface. Ask them what their second-favourite alternative was and why they didn’t pick it.
Anyone who actually researched and wrote the piece can answer in 30 seconds. Anyone who pasted the output of an AI prompt into a document with their name on it will hesitate, hedge, or change the subject. This test costs you one minute and it’s more reliable than any detector you can buy.
This is also why interviews and live writing exercises haven’t gone away. A 20-minute conversation with a candidate about something they claim to have written tells you more than any tool.
A few mistakes I see professionals making, and the better move:
Don’t paste student or contractor work into a detector and treat the score as proof. The false positive rate, especially for non-native English writers, is high enough that you will eventually accuse the wrong person. Use detectors as one signal among several, never as the deciding evidence.
Don’t rely on “AI rewrites it to feel human” tools as a defence. These tools (you can search for them, I won’t link) introduce small grammatical noise to evade detectors. They make writing slightly worse, and the next generation of detectors usually catches them within a few months.
Do build a “human signal” into your own writing process. If you’re a professional whose work might be screened, leave fingerprints in your writing on purpose. A specific anecdote. A real number from your own work. A small admission of uncertainty. These are hard for AI to fake convincingly because they require lived context.
Do calibrate your team. If you manage writers, send them three samples (one human, one AI, one AI-edited-by-human) without telling them which is which. Ask them to identify each. Most teams discover their pattern recognition is much weaker than they thought, and the conversation about why is more useful than any tool training.
The most defensible answer to “was this written by AI?” in 2026 is not a probability score. It’s provenance: verifiable information about where the content came from and how it changed over time. [4]
Microsoft Word, Google Docs, and Notion all keep edit histories. If a 2,000-word document was created in three minutes with no intermediate edits, that’s a provenance signal worth more than any detector score. If the same document has 47 revisions over four hours with text being typed, deleted, restructured, that’s a different signal.
Some publishers are starting to require revision histories with submitted work. Some recruiters are using live coding/writing platforms instead of take-home exercises. Image and video have C2PA content credentials being adopted across major platforms. [5] Text is harder to fingerprint cryptographically, but the workflow signal (revision history) is good enough for most use cases.
If you’re hiring, commissioning, or grading work, ask for the working file with revision history, not the polished output. The fact that you’re asking changes the behaviour. Anyone planning to paste from ChatGPT will think twice.
Three small actions:
1. Pick one writing sample you’re not sure about and run it through the seven-tell checklist. Practice. The pattern recognition gets faster the more you do it.
2. If you manage a team, run the calibration exercise. Three samples, no labels, ask them to score. The conversation is the value.
3. If you’re hiring, change one process: ask for the working document, not just the final draft. Watch what happens to the quality of submissions. The threat of being asked is most of the deterrent.
AI writing isn’t going anywhere. The honest goal isn’t to eliminate it, it’s to know what you’re reading and decide accordingly.
There isn’t a clear answer. Pangram, GPTZero, Originality.ai, and Copyleaks all claim accuracy above 99% on their own benchmarks, but real-world tests vary widely depending on the model the AI was generated with and how the writer prompted it. Treat all detectors as a flag for review, not a verdict.
They keep improving but they’re chasing a target that moves faster than they do. Each new model generation produces text that’s harder to detect, and a one-line prompt change can defeat most detectors. Expect a permanent cat-and-mouse situation rather than a solved problem.
Yes, with effort. Mixing AI drafts with substantial human editing, adding specific personal details, varying sentence structure, and including genuine opinions all reduce detector signal. But the goal of writing isn’t to defeat a detector, it’s to communicate honestly. If you’re using AI, the better question is whether you’re transparent about it.
Depends entirely on the context and the rules in place. AI as a brainstorm partner, an editor, or a research assistant is a productivity tool. AI generating the final draft you submit as your own work in a context that prohibits it (academic submissions, contracted writing) is misrepresentation, not just policy violation.
Be very cautious. The false positive rate, especially for non-native English writers, is high enough that detector-only decisions have led to wrongful accusations. Use detectors to trigger a closer human review and a conversation with the writer, never as the final word.
About this guide
This article was researched and written by Sana Mian, co-founder of Future Factors AI. It draws on independent benchmarks of leading AI detection tools, Stanford research on detector bias, and real-world testing across thousands of student and contractor submissions reviewed through Future Factors training programs.
Sources
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.
