Last updated: October 2026
Teachers use them. Employers use them. Some website owners worry about them. AI detectors promise to tell you whether a text was written by a person or by a machine, and a whole group of websites now sells the opposite: tools that “humanize” AI text so detectors do not catch it. But do any of them really work?
We ran a small hands-on test on October 21, 2026. We took a text written by ChatGPT, a text written by a human in 1891 and the output of a free AI humanizer, and passed all of them through three free AI detectors. The results are not what the marketing suggests.
The Short Answer
- All three detectors flagged the raw ChatGPT text as 100% AI.
- Two of the three also flagged a human-written text as AI, at 41% and 94%. Only one got it right.
- The free humanizer barely helped. One detector dropped to 48%, another stayed at 99% and the third stayed at 100%.
- The humanizer made the text worse, with broken sentences and double periods.
It is one test with a few short samples, so read it as a snapshot and not as a verdict.
How We Ran the Test
We wanted a test that anyone could repeat, so we used only free tools that work without an account. We declined the optional cookies on each site and did not register anywhere.
| Sample | What it is | Length |
|---|---|---|
| AI text | A short essay on the pros and cons of working from home, written by the free version of ChatGPT (first answer, no retries) | 264 words |
| Human text | The opening of the novel A Scandal in Bohemia (1891), a public-domain text from Project Gutenberg | 301 words |
| Humanized text | The ChatGPT essay after passing through the free Basic Humanizer on ZeroGPT’s website | 263 words |
The three detectors were ZeroGPT, GPTZero and Sapling, in their free web versions. We checked that each text was entered exactly as it appears, with no missing letters. We ran each text once.
The Results
| Detector | ChatGPT text | Human text (1891) | Humanized text |
|---|---|---|---|
| ZeroGPT | 100% AI | 40.9% AI | 48.3% AI |
| Sapling | 100% AI | 93.5% AI | 98.9% AI |
| GPTZero | 100% AI | 0% AI (100% human) | 100% AI |
Each detector shows its result in its own way, so we copied the main percentage that each tool reported. ZeroGPT’s headline still said “AI/GPT Generated” for the human text, even with a score of 40.9%.
What We Noticed
1. Raw AI text is easy to catch
The plain ChatGPT essay was caught by all three at 100%. It uses stock phrases such as “In conclusion”, “Moreover” and “Furthermore”, and it keeps a very even rhythm. That is what detectors look for.
2. Human text can be flagged as AI
The surprise was the human text. GPTZero called it entirely human, but ZeroGPT gave it 40.9% AI and Sapling gave it 93.5%. This is a formal, polished piece of writing, and detectors seem to treat that kind of style with suspicion. A classic novel is also a text that is widely available online, which may confuse some tools.
This matters because a detector score is often used as proof. The Mount Royal University advisory on AI detection reports that independent tests found none of fourteen detectors, Turnitin included, reaching 80% overall accuracy. You can read it in their advisory.
3. The free humanizer did not fool them
After humanizing, ZeroGPT went from 100% to 48.3%, which looks like a win until you remember that the same tool gave 40.9% to a real human text. Sapling moved only from 100% to 98.9%, and GPTZero still said 100% AI.
The humanizer’s own page warns that detectors “might still flag” text from its basic public model, and it offers a paid advanced model. We did not test the paid version, so we cannot say how it performs.
4. The humanizer made the writing worse
The rewritten text had problems that a person would not leave in. It used double periods (“focused..” and “every day..”), a sentence fragment (“Spend more time with their family.”), and an awkward phrase (“comfortable, to them which can make them feel calmer”). It also lost some meaning: the original said people may keep working “after their usual working hours”, and the new version says people “might end up working hours”.
In other words, the tool tried to avoid the patterns of AI by breaking the writing, and the result was still flagged.
What Are AI Humanizer Websites?
AI humanizers are online tools that rewrite a text so that it looks less like it came from an AI. Most are free to try with limits and sell a paid plan for better results. Their pitch is simple: paste your AI text, press a button and get something that passes detectors.
Our test suggests being careful with that promise, especially with the free versions. A tool can lower one detector’s score and leave the others unchanged, and it can damage your text in the process. If you are going to rely on one, test it with several detectors, and read the result before you use it.
Should You Use One?
Here is our honest view.
- If a school, university or employer does not allow AI, a humanizer does not make it allowed. Follow the rules and be honest about how you used AI.
- For a website or blog, the real question is quality, not detection. Google’s own guidance says AI can be useful for researching a topic and structuring original work, but that using it to generate many pages without adding value may violate its spam policy on scaled content abuse. You can read it in Google’s page on generative AI content.
- A better humanizer is you. Edit the text yourself, add real examples and experience, and check the facts. That improves the content in a way no automatic rewriter can.
Tips if You Write With AI
- Add your own experience. Specific details, opinions and examples make a text useful, and they are hard for a machine to invent.
- Edit for your own voice. Read it aloud and change what you would never say.
- Do not rely on a detector score as proof. A high score can be wrong, as our test shows.
- Keep your drafts. If your work is ever questioned, notes and earlier versions show how you wrote it.
- Be open about it when it matters. Honesty is simpler than trying to hide AI use.
For writing tools that help with editing and style, see our guide to the best AI writing tools. If you are new to AI, you may also like how one beginner learned from his mistakes when starting with AI.
Frequently Asked Questions
Can AI detectors tell if ChatGPT wrote a text?
In our test, all three caught a plain ChatGPT essay. But they also flagged a human text, so they are not reliable as proof.
Which AI detector was the most accurate in your test?
GPTZero was correct on both the AI text and the human text. We tested one example of each, so we cannot say it is the best in general.
Do AI humanizers work?
The free one we tested lowered one detector’s score and left the other two nearly unchanged, and it made the text worse.
Can a human text be flagged as AI?
Yes. Two of the three detectors flagged our 1891 human text, at 41% and 94%.
Does Google penalize AI content?
Google says that using AI is not the problem by itself. Content made to rank with no value for users can break its spam policies. Quality and usefulness matter most.
Is it okay to use AI for school or work?
It depends on the rules. Check them first and be honest about how you use AI.
Our Take
AI detectors can catch a raw ChatGPT text, but they are not proof of anything, and humanizers are not a reliable way around them. If you write with AI, the best approach is not to hide it but to make the text better: add your own experience, edit it carefully and follow the rules of the place where you publish or hand in your work.
Keep Reading
- Best AI Writing Tools in 2026
- Best ChatGPT Prompts in 2026
- How I Choose Which AI to Use for Each Task
About This Test
This article is independent. We have no sponsorship from any detector, humanizer or AI company, and we do not earn commissions from the tools mentioned. We tested the free web versions in October 2026 with one sample per category, so results may differ with other texts, versions and dates.
Disclaimer: AI products and features change quickly. This article is informational and does not encourage breaking the rules of any school, employer or platform.

