Two tools promise to strip the machine out of machine-written prose. Testing them turned up a bigger problem: the instruments used to grade them cannot tell the difference in the first place.
ZeroGPT rated text written entirely by ChatGPT at 26.5% AI and declared it “likely human written.” Both humanized versions of that same text scored higher.
By the metric everyone optimizes against, humanizing the text made it look more artificial, and the raw model output already passed. Whatever these tools do or fail to do, this detector is not the instrument that can tell you.
The premise
Why does a model trained on human writing sound so unlike a human?
It read everything we wrote. The output still lands with a particular flatness that people now recognize within a sentence or two.
The answer is in the objective. A language model predicts the next token, and it is rewarded for predicting the likeliest one. Averaged across an enormous corpus, the likeliest next token is the one that works in the widest variety of contexts. That is a description of the least surprising possible sentence.
Wikipedia's WikiProject AI Cleanup, which has spent years removing this stuff from articles by hand, puts it plainly:
“LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases.”
Human writing is full of choices that are not the likeliest one. An odd word, an aside, a sentence that runs long because the thought did, a joke that only half lands. Those deviations are the texture. Optimizing them away is not a bug in the training, it is the training working.
Why it matters
Slop has a cost now
Readers got tired first. Search engines followed. Google's guidance targets content produced primarily to game rankings rather than to help anyone, and low-engagement, mass-produced pages are exactly what that describes. Pages that nobody finishes reading do not accumulate the signals that keep them ranking.
For anyone whose revenue depends on organic traffic, that turns a stylistic complaint into a line item. Which is why an entire category of tooling now exists to make machine prose read as though a person wrote it.
Mechanism
How the tools actually work
The two tested here take opposite approaches. One destroys the statistical fingerprint by force. The other edits for the specific habits that give machine prose away.
Approach one
Translation chaining
AI text carries measurable fingerprints: predictable token distributions, uniform sentence lengths, repeating n-grams. Machine translation rewrites text at the grammar level, so word order, morphology and vocabulary all shift. Chain enough hops through distant languages and the fingerprint does not survive.
Passes
4 (2 LLM rewrites, 2 machine translations)
Cost
~$0.0002 and 58 seconds per run
Needs
An LLM key plus two translation services
Sees source
Only at step one
Approach two
Pattern rewriting
Not a library at all, but a 456-line Markdown skill any capable agent can load. It encodes 35 named patterns drawn from Wikipedia's Signs of AI writing, then rewrites against them in a single pass that keeps the original in view the whole time.
Passes
1, plus a self-critique before the final draft
Cost
One agent pass, no API key
Needs
Nothing. No network calls, no dependencies
Sees source
Throughout
The four passes, in order
The chain is genuinely sequential, and each hop is doing something different. Steps one and two are not translations at all, they are rewrites that happen to change language. Steps three and four are pure machine translation.
STEP 01
EN → ZH
LLM rewrite
Instructed to strip the “AI flavor” while translating into Chinese. The last time the chain sees your original.
STEP 02
ZH → JA
LLM rewrite
Same instruction into Japanese, with step one's input and output fed back as few-shot history.
STEP 03
JA → FI
Google Translate
Finnish is agglutinative. Meaning rides in stacked suffixes, not word order, so the structure has to be rebuilt.
STEP 04
FI → EN
Niutrans
A second vendor, deliberately. Reconstructing English from Finnish forces structural change rather than word swaps.
Tinted steps are LLM rewrites; plain steps are machine translation. German and Korean are the other tested intermediate languages.
What the rewrite prompt says
Both LLM steps share one system prompt and one instruction. Translated:
system → 你是一个专业的文案改写专家，精通多语言本地化。
“You are a professional copywriting rewrite expert, skilled in multilingual localization.”
user → 翻译为{语言}，去掉 AI 味道，拟人化改写，只输出结果：
“Translate to {language}, remove the AI flavor, rewrite to sound human, output only the result:”
Note what is absent. No instruction to preserve facts. No reference to the source after step one. No verification pass. The chain is built to defeat a classifier, and it is honest about that.
The test
Finding text that a detector agrees a human wrote
The hardest part of the experiment was establishing a baseline. Detectors flag human writing constantly, so a control was needed: text with a known, verifiable human author. The passage came from a book written before any of this tooling existed, with no model involved at any stage.
Three detectors, four conditions: the human baseline, raw ChatGPT output, and that same ChatGPT text after each humanizer.
Condition
ZeroGPT
Originality.ai
UndetectedGPT
Human baseline
3% AI
under 15%
0% AI
Raw ChatGPT
26.5% called human
exceeds 15% flagged
detected flagged
After Humanify
29.1%
25% or less
0% credits out
After Lynote
31.7%
25% or less
0% partial
ZeroGPT score, by condition
Higher = more “AI”
Human baselineknown author
3.0%
Raw ChatGPTno humanizing
26.5%
After Humanifyhumanized
29.1%
After Lynotehumanized
31.7%
The ordering is backwards. If the score measured what it claims to, humanizing should move text down toward the baseline. Instead both tools moved it up, past the raw model output that ZeroGPT had already waved through as human.
Originality.ai was the only detector that behaved consistently: it flagged the raw ChatGPT text and reported both humanized versions inside its lower band. That is the result the tools are sold on. But its output is a coarse threshold, not a number, so a move from “exceeds 15%” to “25% or less” is a smaller claim than it appears.
What this test cannot support
One text per condition. These are readings, not a sample, and nothing here is statistically significant.
UndetectedGPT ran out of credits mid-test. Its two zeros come from a partially unlocked panel and should not be read as passes.
Originality.ai reports bands rather than values, so its numbers cannot be compared with ZeroGPT's percentages.
Detector models change without notice. These readings describe one afternoon.
The tool labelled Humanify above is recorded as it was reported and has not been confirmed against the pattern-rewriting skill benchmarked in the next section.
The blind spot
What no detector was checking
While the detectors argued about style, a separate test ran on meaning. The same source passage, roughly 1,900 characters of financial writing, went through both tools and every checkable claim was compared against the original.
Fourteen facts: the return figure, the market capitalization, the price target, four named entities, three dates, two more numbers. The pattern-rewriting skill kept all fourteen. The translation chain kept thirteen, and changed three things about what the text asserts.
every millionaire in the world to own one full coinevery billionaire in the world
Factual error. Moves the scarcity argument by three orders of magnitude.
long-term value could rise substantiallylong-term price increases are inevitable
Hedge deleted. A speculative claim became a certainty, in financial copy.
(nothing in the source)“It seems better to be prepared for twists and turns.”
Sentence invented outright.
This is structural, not bad luck. After step one the chain never sees the original again. Each hop rewrites the previous hop's output, so errors compound and no stage has the information needed to notice. By step four the English is four translations from the source.
The pattern-rewriting skill avoids this for one reason: the source stays in context, and its process explicitly asks, before the final draft, whether any fact, name, number, date, quote or citation was added or lost. That question is only answerable if the original is still there.
Both tools cleaned up the surface tells equally well. Both removed every em dash and every curly quote. The difference showed up only where no detector looks.
Conclusions
What holds up
01
Detector scores are not a measurement
ZeroGPT passed raw ChatGPT output and then rated both humanized versions as more artificial. A number that moves the wrong way under a known intervention is not tracking what it claims to track. Do not tune content against it.
02
The detectors disagree with each other
On identical input, one flagged it, one missed it, and one ran out of credits. Any single score is a reading from one instrument on one day, not a verdict.
03
Translation chaining is unsafe for factual copy
It works as advertised on style and it is cheap. It also silently changed a number, removed a hedge, and invented a sentence, in a passage about money. For finance, medicine, law or technical documentation, that cost is not recoverable by a proofread that does not compare against the source.
04
The cheapest tool was the most faithful
The pattern-rewriting approach needs no key, no network and no vendor. It kept every fact and cut nearly 300 characters of filler. Its ceiling is lower, because it edits rather than rebuilds, but it does not trade accuracy for a lower score.
05
Optimizing for the detector is the wrong target anyway
Search engines penalize low-quality, low-engagement content, not text that trips a classifier. Those are different problems. A page that reads as though someone thought about it solves the one that pays.