Can you make AI text undetectable just by rewriting it? We tried two humanizer modes on the same essay. Both changed the writing. Neither changed ZeroGPT's score. Here's what you can learn by reading the three versions.

One library proposal, three versions

On November 18, 2026, we used the OpenAI API to generate a 371-word essay about a fictional library considering longer hours. We ran it through Basic and Stealth once each in Essay mode, without editing or selecting the best attempt. ZeroGPT returned 100% AI for all three versions. Open a passage to read it in full.

Original essay: 100% AI on ZeroGPT

The proposal to keep the branch library open until 8 p.m. on Wednesdays for six weeks should be treated as a trial rather than a permanent change. Because the library has no visitor counts and no outcome data yet, the decision about whether to continue should be based on evidence gathered during the trial period, not on assumptions. At present, the library knows only that current borrowers were surveyed. That information is useful, but it does not tell the full story, because it excludes people who never visit the library at all.

The strongest evidence to justify continuing the extra hour would be clear use of the service by people who are currently underserved by the regular schedule. If the evening opening attracts a meaningful number of visitors each week, especially after normal daytime hours, that would suggest the later closing time fills a real need. The library should also look at whether the people using the extra hour are new users or existing borrowers. If the extra hour mainly serves people who could already come at other times, the case for continuation would be weaker than if it brings in borrowers who otherwise could not visit.

Because the additional hour requires two staff, staffing efficiency matters too. Continued opening would be easier to justify if the number of visitors during the extra hour is steady enough to make the use of two staff worthwhile. The library should also consider whether the extra hour leads to better access to borrowing or other library use, since a later opening is only valuable if it actually helps people use the service. Any evidence from staff observations about demand, patterns of attendance, or unmet need would also be relevant, provided it is collected during the trial.

The survey of current borrowers can support the idea that some users want later hours, but it cannot by itself justify a permanent change. To make a sound decision, the library needs real trial evidence: how many people come, who they are, whether they would otherwise be unable to visit, and whether the extra hour makes good use of two staff. If those results show sustained demand and practical benefit, continuing the Wednesday evening opening would be justified.

Basic rewrite: 100% AI on ZeroGPT

The Wednesday extension should be judged as a six-week trial, not as a permanent shift. The proposal keeps the branch library open until 8 p.m. on Wednesdays, but there are no visitor counts yet and no outcome data yet, so continuation has to rest on evidence gathered during the trial period rather than on assumptions.

What is known at present is narrow: current borrowers were surveyed. That is useful, but it does not capture people who never visit the library at all, and that gap matters when the question is whether the extra hour serves anyone beyond the existing pattern.

The case for keeping it would be strongest if the later opening is clearly used by people the regular schedule does not serve well. A meaningful number of weekly visitors, especially after normal daytime hours, would point to a real need. The library also needs to separate new users from existing borrowers. If the extra hour mainly helps people who could already come at other times, continuation would be harder to defend than if it brings in borrowers who otherwise could not visit.

Staffing puts another condition on the decision. Because the additional hour requires two staff, the number of visitors has to be steady enough to make that use of staff worthwhile. The library should also look at whether the later opening improves access to borrowing or other library use, since the extra hour only matters if it actually helps people use the service. Staff observations about demand, attendance patterns, or unmet need would count too, so long as they are collected during the trial.

The survey alone can support the idea that some users want later hours, but it cannot justify making the change permanent. A sound decision needs trial evidence about how many people come, who they are, whether they would otherwise be unable to visit, and whether the extra hour makes good use of two staff.

Stealth rewrite: 100% AI on ZeroGPT

Two staff will be needed for every extra Wednesday hour, so the six-week plan to keep the branch library open until 8 p.m. should remain a trial rather than become a permanent change immediately. The library has no visitor counts or outcome data yet. Its decision to continue therefore needs to rest on evidence gathered during the trial, not assumptions.

A survey of current borrowers does indicate that some users want later hours. That finding has limits, however: it says nothing about people who never visit the library at all. The more persuasive question is whether the evening opening reaches people underserved by the regular schedule.

Weekly attendance during the added hour will matter, particularly after normal daytime hours. A meaningful number of visitors would suggest that the 8 p.m. closing time meets a real need. The library should also distinguish between new users and existing borrowers. If the hour mainly accommodates people who could already attend at other times, the argument for keeping it is weaker than it would be if it enables visits by borrowers who otherwise could not come.

Staffing efficiency cannot be separated from that assessment. Continued opening would be easier to justify if attendance is steady enough to make the use of two staff worthwhile. The library also needs to know whether the later hour improves access to borrowing or other library use; a later closing has value only if it actually helps people use the service. Staff observations of demand, attendance patterns, and unmet need may also be relevant, provided they are collected during the trial.

The survey alone cannot justify a permanent change. The trial must establish how many people attend, who they are, whether they would otherwise be unable to visit, and whether the added hour makes good use of two staff. Continuing the Wednesday evening opening would be justified only if the results show sustained demand and practical benefit.

What an editor would change

The Stealth version puts the staffing decision first. That's useful: the reader knows immediately why the proposal is a trial. But compare its words does indicate that some users want later hours with the original's can support the idea that some users want later hours. Surveying current borrowers tells us who was asked, not what they answered. Without the answers, keep the claim conditional.

Try this on your own ChatGPT draft: highlight one finding, one proposal, and one uncertainty before you rewrite. Check all three afterward. Use the AI text humanizer to compare drafts, the text detector for a separate estimate, or read the full four-essay test. A detector cannot verify whether a survey finding is true.

1. Specify a Casual, Conversational Tone

Name the reader and the situation before choosing a tone. A customer update, a research discussion, and a personal essay need different voices. Conversational wording can help an informal article, but forcing slang into a formal report makes the writing less useful.

Instead of: "Write an article about climate change"

Try: "Write a casual, conversational blog post about climate change. Write like you're explaining it to a friend over coffee."

2. Ask for Varied Sentence Length

Use sentence length to control emphasis. A short sentence can carry the decision; a longer one can explain its condition. Read a paragraph aloud and notice where several sentences repeat the same opening or rhythm.

Add to your prompt: "Vary sentence structure where it improves clarity. Keep a condition next to the claim it qualifies, and do not split a sentence just to meet a length pattern."

3. Include Personal Touches

A real observation gives a draft something specific to say. Supply an experience you actually had, or label an invented example as hypothetical. Do not ask a model to manufacture a memory and present it as yours.

Try: "Use the observation in my notes. Keep the details I supplied, and do not add a conversation, result, quotation, or personal experience."

4. Avoid "AI Words"

A word is not proof of who wrote a passage. Still, repeated stock transitions can make prose feel interchangeable. Review phrases such as:

  • "Furthermore" and "Moreover"
  • "In conclusion"
  • "It's important to note that"
  • "Delve" and "dive deep"
  • "Landscape" (when not literal)
  • "Leverage" and "utilize"
  • "Comprehensive" and "robust"

Add to your prompt: "Avoid overused transitional phrases like 'furthermore' and 'moreover.' Don't use words like 'delve,' 'landscape,' or 'leverage.'"

5. Keep Voice Without Adding Mistakes

Do not add typos, awkward grammar, or irrelevant tangents to create an impression of authorship. Keep a direct question, a restrained aside, or a sentence beginning with "But" when it belongs in your voice. Correctness and personality can coexist.

Try: "Keep the direct, plainspoken voice of my sample paragraph. Correct errors without making every paragraph sound formal or equally shaped."

6. Specify a Reading Level

Tell the editor what the audience already knows. Keep technical terms that readers need, and explain unfamiliar ones where they first matter.

Options:

  • "Write at an 8th-grade reading level"
  • "Use simple words a high schooler would use"
  • "Write for experts who already understand the basics"

7. Give It a Persona

Describe an editorial approach without inventing credentials. You can ask for a skeptical review or a clear classroom explanation without claiming ten years of reporting or teaching experience.

Try: "Use a skeptical editorial voice: distinguish what the evidence shows from what remains uncertain. Do not add biographical claims about the writer."

8. Request Specific Examples

An example should do a job: demonstrate a choice, clarify a constraint, or show what a reader would change. Use verified source material or a clearly labeled fictional example. An impressive-looking statistic without a checked source is not an improvement.

Add: "Use only the examples and sources I provide. Mark a gap for my review instead of inventing evidence."

9. Break Formatting Conventions

Let the argument decide the structure. A troubleshooting article may start with the failure, a product description with a use case, and an essay with a disagreement. Equal paragraph lengths are not a requirement.

Try: "Don't follow a rigid structure. Some sections can be longer than others. You can have a one-sentence paragraph for emphasis."

10. Edit the Output Yourself

The most effective technique: actually edit what the AI produces.

  • Add your own examples and anecdotes
  • Rewrite sentences in your voice
  • Delete anything that sounds "too AI"
  • Add opinions and personality
  • Include information only you would know

A focused editing pass can make a draft clearer and more specific, but there is no reliable percentage of editing that guarantees any detector result.

Bonus: The Mega-Prompt

Here's a template you can adapt for any content:

Rewrite the full passage for [reader] in a [tone] voice. First identify its argument, facts, examples, and limitations. Reorganize the prose where that makes the argument clearer. Preserve names, numbers, dates, conditions, and uncertainty. Keep fictional examples labeled fictional. Add no facts, sources, anecdotes, or product features. Stay close to the source length. Check the finished rewrite against the original, then return only the revised text. Text: [INSERT TEXT HERE]

A Worked Revision: Keep the Constraint

Fictional source notes: a library tests a desk notebook for four weeks; Maya opens Tuesday and Thursday at 08:30; Leo closes at 18:00; visitor contact details stay in a restricted system; urgent building problems go directly to the duty manager.

Weak draft: The new notebook will transform teamwork and ensure that every issue is resolved quickly. Maya and Leo can use it to share all visitor information and manage urgent problems efficiently.

The weak draft invents a result, removes the schedule, and contradicts two constraints. More fluent wording would not fix it.

Edited example: At the fictional library, a four-week notebook trial will help staff hand over unfinished work. Maya opens on Tuesdays and Thursdays at 08:30; Leo closes at 18:00. The notebook stays at the desk, but visitor contact details remain in the restricted system. Urgent building problems still go straight to the duty manager.

The revision earns its specificity from the supplied notes. It does not promise faster queues or a successful pilot. During a humanizer pass, check the same anchors: who acts, when, what changes, and what must stay outside the tool.

Turn the Tips Into a Review Routine

Keep an untouched source copy, revise one complete passage in the AI Humanizer, and compare the output beside it. Basic and Stealth each support General Writing, Essay, Article, Marketing Material, and Story. Choose the writing type for the actual document, not a score you hope to obtain.

Read the rewrite once for meaning and once for voice. Restore a missing condition before polishing a transition. For a repeatable comparison, use the model-comparison worksheet; for interpreting a scan, read how AI detectors work. Detector results and editorial quality are different checks.

Perfect Your AI Content

Start with a clear prompt, then use AI Undetectable to revise awkward patterns and review the result against your original meaning.

Open AI Humanizer