There's a lot of bad advice floating around about how to make AI text undetectable. People swear by prompt tricks, manual rewriting, running things through QuillBot. We tested seven different approaches methodically — same ChatGPT essay, same detectors, documented results for each. Here's the truth about what works.
| Method | Turnitin | GPTZero | Reliable? |
|---|---|---|---|
| Use a purpose-built AI humanizer (HumanizeTech) | 4–8% | 2–6% | |
| Rewrite from scratch using AI as an outline only | 10–25% | 8–20% | |
| Heavy manual editing (30+ minutes per 500 words) | 20–50% | 25–55% | |
| Asking ChatGPT to rewrite itself more 'human-like' | 65–80% | 70–85% | |
| Running through QuillBot paraphraser | 50–65% | 55–70% | |
| Changing writing style prompts | 75–90% | 80–92% | |
| Adding typos and grammatical errors | 80–90% | 75–88% |
Use a purpose-built AI humanizer (HumanizeTech)
The only method that consistently defeats both Turnitin and GPTZero across all text types. Purpose-built humanizers attack the statistical signals detectors measure — perplexity, burstiness, stylometric consistency — rather than just swapping surface words. HumanizeTech's Academic mode is specifically calibrated for Turnitin. Run your text through, verify the score, done.
Rewrite from scratch using AI as an outline only
Use ChatGPT to generate an outline and key points, then write the actual prose yourself. This works well because the ideas might come from AI but the expression is genuinely human. The downside: it's time-consuming. You're essentially writing the essay yourself, just with better notes. For students who use AI to shortcut writing effort, this defeats the purpose.
Heavy manual editing (30+ minutes per 500 words)
Manually rewriting sentences, changing transitions, varying paragraph lengths. This reduces detection scores meaningfully — but rarely gets below the 20% threshold where you're safe. After 30 minutes of editing, most text still sits at 35–50% AI. You'd need to essentially rewrite every sentence, at which point you're working harder than just writing it yourself.
Asking ChatGPT to rewrite itself more 'human-like'
Prompting ChatGPT to 'write more like a human', 'add imperfections', or 'vary sentence lengths'. This does make marginal improvements but the output is still clearly AI. ChatGPT doesn't know what specific patterns Turnitin or GPTZero look for, so it can't reliably defeat them — it just produces a different style of AI writing.
Running through QuillBot paraphraser
QuillBot's paraphraser does reduce AI scores by swapping vocabulary and rephrasing some sentences. But it doesn't fundamentally change the statistical signature of the text. Turnitin scores typically drop from 90% to 55–65% — better, but still safely in 'flagged' territory. QuillBot wasn't designed for this purpose.
Changing writing style prompts
Using prompts like 'write in an informal tone', 'add personal anecdotes', 'write like a first-year student'. These help shape the output but don't defeat detection. Turnitin's model is trained on informal student writing too — it knows what 'casual academic writing' looks like.
Adding typos and grammatical errors
Deliberately inserting spelling mistakes or grammatical errors. This barely moves the needle on detection scores and actively hurts your grade. Turnitin's AI model is robust to surface noise — it's measuring deep linguistic patterns, not surface form. Don't do this.
The Bottom Line
The only methods that reliably get AI text below the 20% threshold on Turnitin and GPTZero are: writing the text yourself (using AI only for planning), or using a purpose-built humanizer like HumanizeTech.
Everything in between — prompt tricks, QuillBot, manual light editing — gives you a false sense of security. You end up submitting text that still scores 50–70% AI, which is still very much in "flagged" territory.
If you're going to use AI in your writing process, do it right. Use HumanizeTech to finish the job — it takes 10 seconds and costs $4 for a full essay.