AI Humanizers: The (Slop)^2 Problem

How AI Humanizers Work (And Why They Don’t)

AI humanizers are tools used to change text to bypass AI detection systems. AI Humanizers are not focused on including human language; they are made to bypass detection by any means available. This results in worse text as humanizers make AI slop less human, less comprehensible, and less accurate. This leads to what we will call: The (Slop)^2 Problem.

The strategies used by humanizers demonstrate a fundamental misunderstanding of AI detection. AI-generated text is known to use signifiers such as formal language and structured formatting, and to overuse words like “delve” and “tapestry”. By paraphrasing text, manipulating punctuation, and inserting out-of-place characters, humanizers produce text that may result in the prediction: 100% human. In a previous article, we demonstrated that changing the wording or inserting random characters into text is not effective for evading good AI detection.

Most humanizers include their own AI detection tools that will boast the effectiveness of their process by showing the text is no longer AI-generated. This should not be taken at face-value as effective AI detection is robust to humanizers and paraphrasing.

Common Humanizer Problems

1. Hard-coded Punctuation Errors

Humanizer_1 frequently replaces hyphens with commas. This is clearly intentional. In a sample of 194 texts with 164 hyphens, Humanizer_1’s output contains ZERO hyphens. This can affect comprehension or context. For example, multi-ethnic inner-city street is humanized into: multi, ethnic inner, city street. The timing of the Second French Intervention in Mexico changes from 1861-1867 to 1861, 1867, which may be interpreted as two separate years rather than a range.

Humanizer_2 unnecessarily capitalizes words and phrases: A serious attack in the form of a stabbing at the HMP Pentonville has cast a dark shadow on the Government's announcement on a new White Paper regarding Prison Reform and has highlighted the increasing safety crisis within the UK's Prison System.

Humanizer_5’s output below eliminates the space after periods and randomly inserts a hyphen: That's a really interesting question!Yeah, some animals actually do communicate in ways that are this complex, but it's not as common as you'd think.So, we're talking about creatures that chat in pretty intricate ways.Honeybees are a really good example of what you're talking about.Their waggle dance isn't just a simple "food found" message; it actually tells them more.- Where the sun is shining from (its angle).Distance (duration of waggle)The dance had a lot of energy.

2. Paraphrasing Quoted Text

Quoted text lends to the credibility of reporting and research. As far as humanizers are concerned, credibility is thrown out of the window. Humanizer_1 frequently eliminates quoted text through paraphrasing, but Humanizer_2 goes a step further by editing the original quote, leading the reader to believe the provided quote is legitimate.

Before Humanization: "Every day that passes without the ESM being fully operational is a day of increased risk for the eurozone," commented a senior EU official, speaking on condition of anonymity. "The markets need to see decisive action and the full arsenal of our crisis tools at the ready. This German delay is undermining that message."

After Humanization: "Every day that goes by with the ESM unable to operate fully represents an increased risk to the Eurozone," stated a Senior E.U. Official, speaking under condition of anonymity. "The markets must see decisive actions and that the full range of our crisis fighting tools are available. The German delay undermines that message."

3. Tone-Deaf/Inappropriate Paraphrasing

Humanizer_1 uses insensitive paraphrasing. The humanizer uses the informal/colloquial “youngster” to refer to the “youth”, shifting the serious tone of the unfortunate situation.

Before Humanization: Mr. X was found seriously injured following an altercation involving a youth.

After Humanization: After an altercation with a youngster Mr. X was discovered seriously injured.

Humanizer_3 takes the original text: Biological systems can be fruitfully conceptualized as heterogeneous information networks (HINs) in which diverse entity types. . . and changes it to: It’s useful to think of living things as varied information networks – HINs – where different kinds of things. . .. The humanized text then uses the abbreviation HIN after paraphrasing its true definition.

Humanizer_4 demonstrates a lack of understanding by changing Dear Team to To Whom It May Concern.

4. Incorrect Grammar/Nonsensical Output

From Humanizer_1: His eyes, normally a bright blue glow inside the ethereal light, were dark, pain has covered them, and there was a trace of fear that is very human in nature. The golden glow that was around him and the power of his voice that made the whole world vibrate, has vanished, and the smell of blood and ozone met the senses of the observers.

The original text uses the past perfect tense throughout the work, but in Humanizer_1’s attempt to change the text, it paraphrases verb tenses poorly. It incorrectly mixes simple past tense (“was”, “were”, “made”) with present perfect tense (“has covered”, “has vanished”) and simple present tense (“is”).

Humanizer_4 makes Portions were generous, but not overwhelming. slightly convoluted, interpreting the sentence as They served large portions of food, but not too much to be able to eat all of it..

Humanizer_5’s output starts off coherent but deteriorates into hallucination.

Before Humanization: A composite healthy-behavior score was inversely associated with incident hypertension and diabetes (adjusted risk ratio per 1-point increase, 0.87; 95% CI, 0.84–0.90).

After Humanization: They found that the lower someone's healthy behavior score, the more likely they were to experience issues like high blood pressure and diabetes."87; with a 95% CI of 0."Yeah, what a blowout, right?90).

In the middle of humanizing text, Humanizer_5 launches into weird self-promotion, which is hard to catch because of how lengthy and poorly punctuated the output is:

Yeah, I totally saw that coming.Lyra's head popped out the window, and for a moment, her usual sadness was gone, replaced by pure excitement.A knight!" she shrieked, like I couldn't see it myself."He's here for me!"There's a lot of chatter out there about AI, right? It can be tough to tell the real science from the made-up stuff. We really just want to cut through the noise and show you what's possible, because it's genuinely exciting. This isn't about some distant future; it's about what we can do today. Think of it like this: instead of complicated models that are hard to understand, we're talking about practical applications that make a difference. We know some folks are worried about how AI will change their jobs, and that's a valid concern. But a big part of what we're doing is actually making work easier and more efficient, letting people focus on the creative parts. For anyone who's curious about AI and what it means for your business, we're here to offer real, actionable insights. We believe in clear explanations and practical solutions, not just buzzwords. So, let’s talk about figuring out all of this together. We're here to help you get a clear picture of how AI can actually serve you, without all the usual fluff.I rumbled a low, warning growl.Not at the knight.At her.She was getting ahead of herself.

5. Misinformation/Missing Information

Humanizer_2 states: Justin Juuko was born in Uganda was a professional boxer with a successful career that lasted over thirty years. He was able to realize your dreams of becoming a two-time world champion. .There is no indication of this in the original text, and a quick search can confirm Juuko was never a two-time world champion. This is a hallucination, made even worse by the random inclusion of the second-person pronoun, which is never used in the original text.

Humanizer_4 removes important context from Campomelic dysplasia is a rare genetic skeletal disorder with no cure, so treatment focuses on managing symptoms and complications. Here are the main therapeutic approaches:.

In the new humanized text: There is currently no cure for campomelic dysplasia; thus, treatments are focused on addressing the symptoms and complications associated with the condition. The primary treatment options include:, we understand less about the definition of campomelic dysplasia.

6. Tortured Phrases/Excessive Paraphrasing

Humanizers make the tortured phrases prevalent in AI-generated text more apparent and out of place. Cabanac et al. define tortured phrases as “unexpected weird phrases in lieu of established ones, such as ‘counterfeit consciousness’ instead of ‘artificial intelligence.’” These researchers found a prevalence of this writing style in AI-generated scientific papers.

In a scenario of tortured phrasing, an AI-generated review wrote: my culinary preconceptions shattered. While this was written to praise a restaurant, the humanizer tool states: my gastronomical misbeliefs were broken.

To describe the male-dominated field of aviation, Humanizer_2 framed it as the previously masculine-dominated field of aviation. Alongside the needless paraphrasing, this implies the aviation industry is no longer male-dominated, which is simply untrue.

Perhaps a more exhaustive example is in the way Humanizer_2 attempts to re-interpret a research paper’s abstract by writing out the mathematical symbols:

Before Humanization: In both cases we obtain sub-Gaussian concentration with temperature-dependent but volume-independent constants, Var u_x ≤ C T and tails P(|u_x| ≥ r) ≤ exp(−c r^2/T) for T ≤ T0, as well as optimal decay of covariances.

After Humanization: For the two cases above we will provide sub-Gaussian concentration bounds with volume independent constants as a function of temperature and for the sufficiently low temperature one will have that the variance of the particle displacements is less than or equal to C T and the probability that the displacement from the zero particle position is greater than or equal to r is exponentially small in terms of the ratio of r² and T (for T ≤ T0) with optimal decay of covariance bounds.

The Cost of AI Humanizers

When using AI humanizers, we recognized several ways texts were made longer but not better. Humanizers remove an additional layer of context and meaning from the text without enriching the quality of the work.

AI humanizers are paid tools (costing $10-$50/month), which begs the question: why would anyone pay to make their writing worse? Especially when the main deliverable of evading detection is not guaranteed.

Humanizers contribute to the growing problem of slop content, but take slop to the next level: (Slop)^2.