Can You Use ChatGPT Prompts to Avoid AI Detection Tools? Here's What Happens

Straight answer: yes, certain ChatGPT prompts will lower your score on tools like GPTZero, Originality.ai and Copyleaks, but the scores themselves are so unreliable that "beating" them proves nothing and can quietly wreck the quality of your writing. I've run the same piece of content through three detectors and got three different verdicts. Chasing a number on a screen is the wrong goal; writing something that sounds like you is the right one.

What these detection tools are really measuring

AI detectors don't read for meaning. They score two things: perplexity (how predictable your word choices are) and burstiness (how much your sentence lengths vary). Human writing tends to be messy, we mix short punchy sentences with long rambling ones, we repeat words we like, we go off on tangents. GPT-4 and GPT-5 style output, left to its own devices, is smooth and even. Every sentence sits around the same length. That evenness is what gets flagged, not "AI-ness" in any meaningful sense.

Which means the tools aren't detecting AI. They're detecting a writing style that happens to correlate with AI output about 70 to 85 percent of the time, depending on which vendor you believe and which day you ask them. That gap matters more than most articles on this topic admit.

The prompts everyone tries (and what they do)

If you've searched this topic, you've seen the same prompt formulas everywhere:

• "Write this in a more human, conversational tone, vary sentence length, use contractions"
• "Add small imperfections and casual asides like a real person writing"
• "Avoid common AI phrases, don't use words like delve, moreover, in today's world"
• "Write at an 8th grade reading level with some grammatical looseness"

These do change the output. I've tested them, not just once. I took a 600-word blog intro I'd written with ChatGPT for a client in the coaching space, ran it through Originality.ai as-is, then ran three "humanised" versions through the same tool an hour later.

The original scored 94% likely AI. After one round of "make this sound more human and vary the sentence length," it dropped to 61%. After a second pass asking it to "write like a slightly tired person typing quickly, with some sentence fragments," it dropped to 22%. I ran the exact same 22% version through the tool again fifteen minutes later, changing nothing, and it came back at 38%. Same text, different score, same tool. That inconsistency alone should tell you something about how much weight to put on the number.

A real client moment that changed how I think about this

Last year a small ecommerce client asked me to check whether the product descriptions her VA had written with ChatGPT would "pass" before she uploaded them to her site. I ran them through GPTZero. Six out of ten flagged as AI-generated. So we ran them through a "humanise this" prompt loop and got the score down. Then, out of curiosity, I ran a paragraph she'd typed herself, no AI involved at all, straight from an email she'd sent me the week before. GPTZero flagged it as 71% likely AI-generated. Her own words, typed by her own hand, called out as a machine.

That's the moment I stopped treating these scores as facts. They're guesses, dressed up as measurements. OpenAI retired its own AI text classifier in 2023 after admitting it correctly identified AI-written text only about 26% of the time. If the company that built the models can't build a reliable detector for its own output, a browser extension charging £10 a month isn't going to solve it either.

So what does move the needle

If you want content that reads as human because it is more human, rather than gaming a flawed scoring system, here's what worked across dozens of pieces I've edited:

• Add one real detail the model couldn't know. A specific number, a client name (anonymised if needed), a date, a mistake you made. AI can't invent your Tuesday afternoon in 2019 when a client cancelled on you five minutes before a pitch.
• Cut every third linking phrase. "Furthermore," "in addition," "it's worth noting" - these are the tells. Human writers jump between ideas without announcing the jump.
• Vary paragraph length on purpose. One sentence. Then four. AI defaults to a rhythm; break it manually.
• Read it out loud. If you wouldn't say it to a friend over coffee, rewrite it. This single step catches more "AI voice" than any prompt does.
• Keep your own opinions in, even the blunt ones. Models are trained to hedge. People aren't. "I think this is a waste of money" reads as human because most AI output is trained away from saying exactly that.

I use a version of this checklist whenever I'm working through prompts that help write better business content, and the improvement isn't in the detector score, it's in the engagement. Content that sounds like a person gets more comments, more shares, more replies. That's the metric that pays your rent, not a percentage on a browser plugin.

The uncomfortable bit nobody wants to say out loud

Here's the part that annoys people when I say it in webinars: most businesses aren't worried about "AI detection" in any technical sense. They're worried about getting caught sounding lazy. Those are different problems with different fixes. A "humanise this text" prompt fixes the first worry and does nothing for the second. I've seen "humanised" AI copy that still says nothing, still has no point of view, still could have been written about any business in any industry. Changing the sentence rhythm doesn't give it substance.

The businesses getting hurt right now aren't the ones whose content scores 80% on a detector. They're the ones whose emails, blogs, and product pages have started to sound identical to every competitor's, because everyone is running the same prompts through the same model with the same defaults. I wrote about this in why every small business email sounds the same now, and the fix there costs nothing, it's the same fix that works for detection scores: put in something only you know.

There's a knock-on effect too. Marketers obsessing over detection scores are missing that open rates are already dropping across the board because inboxes are full of smooth, samey AI copy that readers have learned to skim past without reading. I've broken down the numbers behind that in why AI is quietly wrecking email open rates, and it's a bigger business risk than any detector flag.

I've broken down ChatGPT usage limits for 2026 in plain English.

Where this matters versus where it doesn't

Context changes everything here. If you're a student worried about Turnitin, that's a different conversation with different stakes and I'd steer you toward writing your own draft and using AI to edit, not generate. If you're a job seeker worried that ATS software or a recruiter will flag your CV as AI-written, I'd focus less on gaming a detector and more on making the document represent you. My piece on using ChatGPT to write a resume that gets interviews covers how to keep your actual achievements and voice in there rather than letting the model flatten everything into generic phrasing, which incidentally is also what gets CVs binned by human recruiters, detector or no detector.

If you're a business owner publishing blog posts or web copy, almost nobody checks these with a detector at all. Google has said publicly it doesn't penalise content for being AI-written, it penalises content for being low quality, however it was produced. So the entire premise of "will this get flagged" is often solving a problem that doesn't exist for your situation. The real risk is publishing something dull, not something detectable.

A short, practical process if you do need to reduce a detector score

If you have a genuine reason (a client contract that specifies "human-written," a platform policy, a college requirement), here's the sequence that holds up better than a single clever prompt:

1. Write your own rough draft first, even if it's messy bullet points, so there's real thinking underneath before AI touches it.
2. Use ChatGPT to expand or restructure, not to originate the whole thing from a one-line brief.
3. Go back through and manually rewrite the first and last sentence of every paragraph in your own words. These are the sentences detectors weight most heavily and the ones AI defaults hardest on.
4. Add one concrete, unverifiable-by-AI detail per section, a date, a number, a name, an opinion.
5. Read it aloud once. Anything that sounds like a LinkedIn post, cut or rewrite.
6. Check on one detector only, not three. Checking on multiple tools just gives you three conflicting numbers and no clearer answer, as I found out firsthand.

This takes maybe 20 extra minutes on a 1,000-word piece. It also happens to be exactly the process that produces better content anyway, which is the bit that should tell you the detector was never really the point. For businesses trying to build this into their everyday workflow rather than one document at a time, this is usually where working with someone as an AI consultant for small business earns its cost, because the fix isn't a prompt, it's a repeatable process your whole team uses.

What I'd tell you if we were having a coffee about this

Stop optimising for a machine that can't reliably tell the difference between your own handwriting and a chatbot's. Optimise for a reader who can absolutely tell the difference between something written with care and something bulk-produced to hit a word count. That reader is the one deciding whether to buy from you, reply to your email, or hire you. The detector never was. Keep an eye on the broader shifts too, I round up what's changing in tools and policy each week in my AI news roundup for small business, because the rules around AI content and disclosure are still moving and worth watching even if the detectors themselves stay unreliable.

Related reading: fun prompts to ask chatgpt.

Related reading: funny chat gpt prompts.

Related reading: ai photo editing strengths.

If your question is a different ChatGPT prompts one, the ChatGPT prompts guide lists every answer I have written.

If you want your own numbers, use the free tool that writes the prompt for you.

Frequently asked questions

Can a specific ChatGPT prompt guarantee my content passes an AI detector?

No. Prompts asking for varied sentence length, contractions, and casual tone reliably lower detector scores, but the same piece of text can score differently on different tools, or even the same tool at different times, so there's no guarantee, only a probability shift.

Do AI detectors give false positives on human writing?

Yes, regularly. I've had human-typed text score over 70% likely AI on GPTZero, and OpenAI itself retired its own detection classifier in 2023 after it correctly flagged AI text only around 26% of the time.

Will Google penalise my blog if it scores high on an AI detector?

Google has stated its ranking systems reward quality content regardless of how it was produced, so a high AI-detector score alone doesn't cause a ranking penalty; thin, generic, low-value content does, whether a human or a model wrote it.

What's the better alternative to gaming an AI detector?

Write your own rough draft first, use ChatGPT to expand or tighten it, then manually add specific details, opinions, and examples the model couldn't have generated. This produces better content and, as a side effect, usually lowers detector scores too.

Further reading

• OpenAI's own platform documentation
• Anthropic's own documentation