How to Tweak Undetectable AI To Make Its Output Better

Can humanized text really bypass AI detection?

Undetectable.ai can quickly “humanize” AI-generated content by modifying, minimizing, or completely removing AI identifiers from the text to avoid triggering AI detectors. Unfortunately, things like structure and readability are often lost in the process.

If used responsibly, this feature can greatly improve content creation and distribution. But the technology is far from perfect. To conserve both humanness and quality, you’ll need to edit Undetectable.ai’s humanized output.

What Humanized Text Looks Like

First, let me set the record straight. 

“Humanized text” is different from “human text” or “human-written text.” The former refers to content that was originally generated by AI and subsequently rewritten by AI, all to help it pass AI detection.The latter refers to content that was originally authored by an actual human being.

Take this article’s introduction. This is an example of human-written text (I promise; I wrote the whole thing from scratch).

Now, if I asked a generative AI platform (like Claude AI or ChatGPT) to write an introduction for this article and then ran its response through Undetectable.ai’s Humanizer, the result would be humanized text.

(You’ll see examples of these later on).

How to Tweak Humanized Text

Ideally, you’d go in and manually edit the content yourself. 

But if you’ve got dozens of texts to humanize and not a lot of time to do so, it can be difficult pinpointing the areas that need editing. And if your output is particularly erroneous (for the more aggressive AI detectors), manual editing could take a while. So going in without a game plan can be inefficient.

With that being said, here’s your five-step game plan for tweaking humanized text.

Step 1. Experiment with Output

The quickest and easiest way to tweak Undetectable.ai’s humanized text is to experiment with the output parameters.

You’ll notice on Undetectable.ai’s platform that the interface has several buttons and drop-down menus around the textbox. Exhibit A:

These options are split into three different categories: readability level, purpose, and human level. Each one modifies and refines how the humanized output will read in subtle but noticeable ways.

For example, text with High School readability will most likely be simple and straightforward. Whereas text with Doctorate readability will read like a dissertation or scientific journal. 

There are actually quite a few parameters in each category. So it’d be good to familiarize yourself with all the parameters, if possible. 

Here’s a quick example to illustrate how the modifiers work. I asked ChatGPT to explain how to make ice cream at home.

Then I created two humanized versions using Undetectable.ai. The only thing I changed were the output parameters.

Readability: High School

Purpose: General Writing

Human Level: More Readable

Readability: Marketing

Purpose: Essay

Human Level: More Human

Same platform, different rules.

These are what you’d call humanized text.

If they don’t read right the first time, you can always go in and play around with the modifiers until you’re satisfied with the results.

Step 2. Use AI Detectors to Find Problem Areas

Some AI content detectors like Copyleak, ZeroGPT, and GPTZero will actually highlight which sentences or phrases seem AI. Like so:

If you know which sentences are problematic, you can go in and rewrite the sentences yourself. 

For this example, ZeroGPT says the first two paragraphs are AI. So I’ll focus on rewriting the first one to improve its AI score.

Not the quickest or the easiest option, but it ensures quality and readability. And if your output scores better than this one (say, 30% - 60% AI), it’ll be easier to find and modify offending sentences.

Step 3. Change Sentence Structure & Syntax

One of the biggest tells of AI text is “burstiness” or sentence structure. 

You’ve probably realized by now that AI loves long, uniform sentences. Take the “homemade ice cream” example. The first three sentences in the example measure 8, 22, and 23 words long, respectively. The second paragraph is 7, 24, and 18.

But if you humanize it, the sentence lengths are more varied: 8, 3, 20, 9, 15.

That’s because AI tends to write following a pattern. Humans, on the other hand, are more unpredictable. Some love long sentences and others prefer short phrases. But, more often than not, we write with a mix of both.

So if an AI generator gives you something structured and monotone, try to give it more rhythm and burstiness by doing the following:

Vary the sentence lengths. If you start with a 25-word sentence, follow it up with two short sentences (3-10 words). After a while, stick to comfortably medium-length sentences (5-15 words). Then throw in a sudden 30-word monster–just to keep things interesting. Remember; variety is the spice of life.

Humanized Text: Indulging in a scoop of ice cream on a scorching summer day is a simple yet deeply satisfying experience. The immediate sensation of coolness, as the creamy icy treat touches your taste buds is truly unmatched. With each spoonful it feels like time slows down allowing you to fully appreciate the moment and escape from the heat. 

Modified: A scoop of ice cream on a scorching summer day. Sweet, simple, and oh-so satisfying. The immediate sensation of sweet, creamy coolness as the icy treat touches your taste buds is truly unmatched. Time slows with each spoonful. Each bite, a welcome oasis in the dusty desert sand.

Don’t be afraid of conjunctions. Words like “and,” “so, “but,” “because,” and “yet” can keep your sentences flowing together. You can even use them at the start of the sentences for a more modern, conversational tone.

Humanized Text: Indulging in a scoop of ice cream on a scorching summer day is a simple yet deeply satisfying experience. The immediate sensation of coolness, as the creamy icy treat touches your taste buds is truly unmatched. With each spoonful it feels like time slows down allowing you to fully appreciate the moment and escape from the heat. 

Modified: Indulging in a scoop of ice cream on a scorching summer day is a sweet, simple treat. And yet it’s so deeply satisfying. The immediate sensation of coolness as the creamy, icy treat touches your taste buds is unmatched. And with each spoonful, it feels like time slows down. The dry, scorching heat may surround you. But with the ice cream in your hand, you are unbothered. Invincible.

Experiment with passive/active voices. Sometimes rearranging the words or phrases in a sentence can greatly improve its readability and appeal. 

Humanized Text: Indulging in a scoop of ice cream on a scorching summer day is a simple yet deeply satisfying experience. The immediate sensation of coolness, as the creamy icy treat touches your taste buds is truly unmatched. With each spoonful it feels like time slows down allowing you to fully appreciate the moment and escape from the heat. 

Modified: Indulging in a scoop of ice cream on a scorching summer day is a simple yet deeply satisfying experience. When the creamy, icy treat first touches your taste buds, the sensation is like no other. It is well and truly unmatched. You fully appreciate each spoonful. Between the rich, creamy texture and the biting cold, time slows, and you’ve escaped the heat.

Step 4. Play with Vocabulary

AI humanizers also use the same lofty, flowery, or (borderline) pretentious vocabulary that AI generators use from time to time. This doesn’t happen often enough but when it does, it’s definitely noticeable.

Let’s take that same humanized ice cream text again. Here’s what GPTZero has to say about it:

Terms like “scorching” or “deeply” (to describe something) aren’t quite as conversational as other words. And the first few sentences, while understandable, are still quite wordy. 

In cases like these, you can try zeroing in on the vocabulary. Swap chosen words out for more natural-sounding phrases, or try a mix of simple words and complex synonyms. 

“Indulging in a scoop of ice cream on a scorching summer day is a simple yet deeply satisfying experience.”

This could be rewritten as:

“A scoop of ice cream on a hot summer day is a simple yet indulgent experience.” 

It’s shorter, sweeter, and much easier to read. 

And even though GPTZero still has a problem with it, ZeroGPT says it’s fine. 

Step 5. AI + Human Editing

While I don’t recommend using this option too often (I’ll explain why in a later section), it can give you more workable results from time to time. 

Say your humanized text isn’t reading right or is still pinging the radar of tougher AI detectors. You can ask a different AI generator to rewrite the text so that it “reads better” or “sounds more human.” 

Then you run the new output through Undetectable.ai’s humanizer again. If you do this a few times (maybe 1-3 tries), you should ideally end up with content that does the following:

1) retains the message of your original text, 

2) passes as 100% human on AI detectors, and 

3) is easy to read.

Running the output through ZeroGPT reveals an 81.6% ping for AI content.

So I have Undetectable.ai humanize ChatGPT’s original text. Once it’s humanized, I check the score.

Better, but not quite 0% AI (which, in this example, is the score I want). So I copy the humanized result, head on over to Claude, and ask Claude to rewrite the new text so that it “reads better.”

I run Claude’s output through Undetectable.ai’s humanizer one more time:

And give it to ZeroGPT for the final verdict.

This method can be a bit of a back-and-forth. I actually had to ask Claude to rewrite the output so that it “seems more human” twice. Then I had to humanize the final result for it to score that perfect 0%.

But if you’re trying to ensure your content doesn’t register as AI on any detector, rewriting it with the help of AI is a viable and efficient option. 

Just be sure to do some quick, cursory editing before posting the final take to ensure quality and readability. If you read the 0% AI sample version again, you’ll see a few grammatical and typographical errors sprinkled throughout. Temper AI editing with human editing for best results.

Why Do We Need to Tweak Humanized Text?

AI humanizers will sometimes rewrite generated content so that it’s less uniform and more flawed. 

The concept is understandable. Humans make mistakes when writing, and sometimes, those mistakes go unnoticed. Humans also have a unique voice and style when writing, which may or may not follow technical grammar, syntax, or vocabulary rules. 

Poetic license, you know.

So, to emulate this, humanizing technology will add these attempts at “nuances”..” Unfortunately, they can negatively impact the output quality. 

Hence the need to edit humanized text. 

If you’re putting content out there, it’s not enough to just have it pass AI detection. It’s meant to be consumed by humans, so–it needs to provide value. If you keep the awkward phrasing and typographical errors, chances are no one’s going to consider your content credible.

Tweaking humanized text ensures optimal quality and readability. It may even improve the text’s ‘perplexity’ and ‘burstiness’ scores by adding the human touch.

Do’s and Don’ts of Humanizing AI Text

There’s a right and wrong way to humanize AI text, and doing it in a questionable way can negatively affect your content. It’s possible to “over-humanize” content to the point that it no longer reads naturally.

That being said, let’s examine some practical guidelines for humanizing AI content. These guidelines help dictate Undetectable.ai’s humanization process, giving us further insight into how the program modifies AI text.

• Do NOT Add Grammar/Spelling Mistakes

Back then, adding grammatical and typographical errors to text was a foolproof “bypass AI detection” plan. Content that had perfect spelling and grammar all throughout could be considered too perfect–and thus most likely written by AI. A couple of misspelled words, some questionable grammar, and poor punctuation usage were enough to fool earlier detectors.

Thankfully, technology has caught up. Though it still needs plenty of refining, hiding behind a few typos won’t cut it anymore. 

Regardless, deliberately adding grammar and spelling mistakes isn’t an advisable method. It jeopardizes the quality - and, by extension, the credibility - of your work. The only goal it achieves is fooling AI. It can no longer provide any sort of value to actual human users.

So if you notice errors in humanized text (like the ones in the samples I shared earlier), sort them quickly or take them out.

• Do NOT Ask AI to “Sound More Human” (too often)

If you remember, this is what we did in Step 5: Put in a prompt, wait for the program’s output, and then ask the program to “sound more human.” You can also copy the output from one AI generator and ask another AI generator to humanize it.

Here’s where I’ll explain why you shouldn’t do this all the time.

Remember that generative AI has limits. It knows, in theory, what it means to “sound human.” But it doesn’t have the intrinsic analyze-and-rewrite process that AI humanizers do. It is not programmed to consider burstiness, perplexity, and redundancy. It will most likely revert to using conversational vocabulary and slang terms (when appropriate), and the result won’t feel as human or as natural as humanized text.

This is known as optimizing AI content. If it’s done too many times, the output may end up with an awkward tone and low readability score.

Here’s what happened when I tried to over-optimize ChatGPT’s response to the prompt, “Can meditation help someone with anxiety?”

The original response:

And after I asked ChatGPT to “sound more human” 5 times:

We went from a generic Wikipedia article to your friend’s mom trying to convince you she’s “the cool mom.” Definitely more conversational but also a little more awkward. 

If that makes sense.

• Do Follow the E-E-A-T Principle

E-E-A-T stands for Experience, Expertise, Authoritativeness, and Trustworthiness. According to experts, this principle is one of the best SEO practices you can observe. 

People want to read content from experts who speak from experience and with authority. This makes them trustworthy–a huge selling point for the average consumer.

AI-generated content often lacks one or more of these factors (especially experience and authority). So by modifying the text so that it fulfills all four requirements of the E-E-A-T Principle, AI detectors will be more inclined to assume it was written by a human. 

And That’s a Wrap

Undetectable.ai’s ability to humanize AI-generated text does not mean humans are no longer needed in this regard. If anything, it’s just further proof that humans and AI can (and should) work together to create optimal content. The humanizer handles the heavy lifting. We humans take care of the finer details.