I started with a piece I was happy with. Researched using AI, notes and thoughts original and drafted with the help of AI (I hate a blank screen, so this process helps me overcome 'writer's block'), then edited by AI as I'm not great at writing with 'active voice' as much of my writing has been business oriented. I then ran it through Grammarly, as I want to learn to write better (failed English at school), and got a 97% grammatically correct score. Happy to publish as far as I was concerned.
Then it went pear-shaped!! I thought I'd check how I was tracking through the eyes of the AI masters and get a score. So I ran it through Originality.AI and it came back flagged as mainly AI output. Crap. I write a lot of stuff for many people (I'm a ghostwriter, amongst other things), so it is important that what I write reads as 'original', and this is where it got interesting and sucked me down yet another AI rabbit hole for a whole weekend. When I emerged, I had a different view on what was going on.
I then put the article through several well-regarded 'humaniser' AIs. The next step was to use an AI detector and try what is apparently the gold-standard judge, Originality.ai. After humanising my apparently non-human writing, I got a gold star. Great. But not great. What got a pass was a seriously crap piece of work, and even though it was deemed human, in my eyes was unreadable.
So I tried again and put the humanised version through Grammarly, which graded it as poorly written and gave me a 70% score. So I fixed the grammar and got it back up to 95%. Back to Originality. Then it was flagged again as 100% AI-generated. I went through this process using various detectors and humanisers to try and crack the code, and ultimately gave up. Annoyingly, I took pieces that passed cleanly as human output, changed a few commas and swapped two words, and watched Originality score the virtually identical text as 100% AI.
I went in assuming the detectors knew something I didn't. They don't. I'd rather write something legible that reflects what's in my head, even if it gets flagged as AI, than produce crap that doesn't get my point across.
I also ran pieces from The Guardian and The Conversation through Originality. They passed. But the writers there have editors with skills the rest of us don't, so that doesn't tell me much. I'm giving up. If my writing gets flagged as AI, so be it.
There's nothing wrong with using AI to get a point across, or to put into words those darting thoughts we couldn't get down before. If the process is AI-assisted, what's the harm? Use AI to write better. The line is whether you're churning out slop and getting lazy. As long as you're not, you're using the tool as it's meant to be used.
What the detectors measure
The main tools score two things. The first is 'perplexity', how predictable the text is to a language model. If the model could have guessed the next word, perplexity is low. The second is 'burstiness', the variation in sentence length and complexity that human writing tends to carry and machine writing tends to flatten.
Low perplexity and low burstiness read as AI. The problem is what else reads that way. Polished prose has low perplexity by construction. When you write well and edit until the seams are gone, you produce exactly the signal the detector is trained to distrust. The tool is not detecting authorship. It is calling polish a machine.
OpenAI retired its own detector
The clearest admission came from the company with the most to gain from a working detector. OpenAI launched its AI Text Classifier on 31 January 2023 and withdrew it on 20 July 2023, citing low accuracy. On its own test set of English text, the classifier correctly identified 26% of AI-written text as likely AI. It also flagged human-written text as AI 9% of the time, and was unreliable on any document under 1,000 characters.
A tool that catches a quarter of what it is looking for and wrongly accuses nearly one in ten humans is not a tool. OpenAI trains the models that produce the text, could not build a reliable detector for their output, and stopped pretending it could.
The bias built into the method
Perplexity does not fail at random. In a 2023 paper in Patterns, Weixin Liang, James Zou and colleagues at Stanford ran seven GPT detectors over a set of essays. The detectors flagged 61.3% of essays written by non-native English speakers, drawn from a TOEFL corpus, as AI-generated. They flagged essays by native speakers about 5% of the time. More than half the non-native writers were misclassified.
Perplexity rises with vocabulary range and syntactic complexity. Non-native writers use simpler words and shorter sentences, so they score lower. Their text falls within the band that the detector flags as machine-generated. The detector reads a smaller vocabulary and simpler sentences and calls it a machine.
Zou's recommendation was to avoid these detectors where the stakes are real: job applications, admissions and assessment. The method has a demographic baked in. The bias punishes humans, not machines.
Turnitin's number
Turnitin introduced AI detection to educators in April 2023, with a stated false-positive rate under 1% at the document level. By June, those numbers were already outdated: the real-world document-level rate was higher, and the sentence-level rate was around 4%. The tool had particular trouble with mixed AI and human prose: 54% of false positives sat next to AI-written sentences. Turnitin added an asterisk to any score below 20% to indicate it is less reliable.
Vanderbilt disabled the feature in August 2023. The university had submitted 75,000 papers to Turnitin in 2022. At a 1% false positive rate, that's roughly 750 papers wrongly flagged, on the vendor's own best-case number, before you account for the higher real-world rate or the non-native bias. Vanderbilt also noted that it had no visibility into how the tool arrived at its scores.
The paraphrase attack
The detectors don't even hold up against a determined attempt to beat them. In 'Can AI-generated text be reliably detected', Vinu Sadasivan and colleagues at Maryland ran a recursive paraphrasing attack against the full range of methods: watermarking, neural network classifiers, zero-shot detectors and retrieval-based tools. A light paraphraser applied on top of the model broke all of them. On 300-token watermarked text, recursive paraphrasing dropped the detection rate from around 99% to 9.7%.
There's a theoretical ceiling on detection. As language models get better at sounding human, AI and human writing blur, and even the best detector trends toward a coin flip. Detection isn't an engineering problem waiting for a cleverer tool. The cap is set by how human the AI sounds, and that bar keeps rising.
The humaniser trade
Which brings me back to the loop. The only way I found to reliably pass the detectors was to make the writing worse. Shorter sentences, plainer words, some grammatical roughness left in on purpose. The humaniser tools automate exactly that, and Grammarly's objection was correct: the humanised text was worse, because degrading it was the point. You are not removing a machine fingerprint. You are adding enough noise that a noise detector relaxes.
That is a poor trade in any professional context. The case for using AI is to produce better work faster. If the price of passing a detector is writing you would not have published on your own standards; you have used two tools to reduce your work to a standard below where you started. The detector wins the narrow game. But you lose the one that matters.
The reader is the better detector
Here is the part that reframed it for me. The most reliable detector of AI text is not a tool; it is a person who uses these models every day. In a 2025 paper, Jenna Russell, Marzena Karpinska and Mohit Iyyer tested frequent LLM users against AI-generated text. A majority vote of five readers misclassified one article out of 300, and held up even against text that had been paraphrased or humanised to evade machines. That beats the commercial detectors!
So the audience you are trying to slip past with a humaniser is already better at this than the software. If your readers use AI, and likely in most professional fields they will be, they are reading for the tics the detectors cannot reliably catch, and the humanisers cannot remove without wrecking the sentence. The gremlins that give it away are not in the perplexity score. They are in the 'not X but Y' constructions, the significance language, the tidy three-part lists, the em dashes. A human reader intuits that, but a machine is oblivious to the fact that it's slop.
Where I ended up
The detector won in the end: I couldn't beat it without reducing my writing to absolute crap, so I gave up trying not to write like an AI. Mind you, on that note, maybe because I'm firmly on the 'spectrum', I already write like an AI!! I often wonder about this, as my preference is mostly to deal with AIs rather than many humans.
Maybe that is the best outcome, but for the wrong reason. The tools do not measure what they claim; they carry a bias they cannot see; they fold under a paraphrase; and their own makers have walked away from them. Treating them as arbiters of authorship is a category error.
The test that holds is the old one. Would this pass a reader who knows the subject and uses the same tools you do? If it does, the detector's verdict is noise. If it doesn't, no amount of humanising will save it, and the best fix is to write it better.