How AI Detectors Work and How to Humanize Your AI-Written Text

As the use of AI tools like ChatGPT becomes more common in writing, so does the use of AI detectors. If you’re using AI to support your writing, you’ll want to understand how these detectors work to avoid being flagged for AI-generated content. This guide breaks down the key elements that AI detectors analyze and provides a step-by-step example of how to manually “humanize” your text, making it undetectable while improving its quality—all without relying on ineffective AI humanizer tools.

[00:15.932]

Many writers are surprised when their text, even after some editing, is flagged as 100% AI-generated. This happens because AI detectors look for subtle patterns that go far beyond simple word choice. To effectively edit AI content, you need to understand what these patterns are and how to manually restructure your writing to remove them.

[00:43.085]

The goal is to transform a paragraph that is easily identified as AI into one that passes as human-written. This isn’t about tricking the system but about taking control of the content and infusing it with human nuance and style. Let’s first explore how these detectors operate.

How Does AI Detection Work?

[01:02.830]

Two of the most important statistical metrics AI detectors use are perplexity and burstiness. Perplexity measures how “surprised” a language model is by a piece of text. Human writing is often varied and unpredictable, leading to higher perplexity. In contrast, AI-generated text tends to be more statistically “expected,” following predictable patterns, which results in low perplexity. Burstiness refers to the variation in sentence length, structure, and word choice. Humans naturally write with fluctuations—mixing long, complex sentences with short, punchy ones. AI, on the other hand, tends to produce text that is more uniform and consistent, lacking this natural rhythm.

Human writing tends to have higher perplexity because it’s more varied and unpredictable. AI-generated text is typically more statistically “expected” and has low perplexity.

[01:40.846]

AI detectors also analyze sentence structure and grammar, specifically looking for unnatural consistency and polish. AI-generated text is often grammatically flawless and stylistically consistent—sometimes too consistent. This perfection can be a red flag. Furthermore, AIs often fall into patterns of repetition, using the same words, phrases, or sentence structures in subtle ways. Another key indicator is the formulaic transition usage. AIs frequently overuse logical connectors like “therefore,” “however,” and “in conclusion” in a robotic manner, rather than integrating them naturally into the flow of the text.

[02:19.462]

The vocabulary and word choice of an AI are also scrutinized. AIs tend to use a “safe vocabulary,” avoiding rare or colloquial words unless specifically prompted. This makes the language feel generic. Because AI models are trained to predict the next most likely word, their output follows high-probability paths, leading to predictable word sequences. Human writers are more likely to make surprising or creative word choices that deviate from these predictable patterns. For academic writing, while avoiding colloquialisms is expected, the overall predictability can still be a tell-tale sign of AI generation.

[03:17.915]

Finally, detectors assess the depth and coherence of the writing. AI-generated text often exhibits surface-level coherence, meaning it makes sense on the surface but may lack deep insight, originality, or nuanced argumentation. It can also be padded with filler phrases—generic statements that sound meaningful but don’t add real value to the text, such as “It is important to consider all perspectives…” These phrases are often used to increase word count without adding substance, another pattern that detection tools can pick up on.

How to Humanize AI-Written Text

[04:57.485]

One of the most effective ways to humanize AI text is to introduce intellectual hesitation. AI tends to write in absolutes, presenting information as factual statements. In academic writing, scholars rarely do this. Instead, they hedge their claims. Use cautious language to show that you are interpreting information rather than stating undeniable facts. Instead of saying “This proves X,” try “This may suggest X,” or “It appears that…” Using words like can, may, suggest, seem, believe, and likely introduces a layer of human analysis and nuance that AI often lacks.

[07:09.117]

A critical step is to vary sentence openings and syntax. AI often starts sentences in the same repetitive way (e.g., “This study shows…”, “It is important…”). To break this pattern, restructure your sentences. Start with an introductory or dependent clause, or use an inverted structure. For example, instead of starting with the subject, begin with a different part of the sentence. This creates the “burstiness” and unpredictability characteristic of human writing and forces you to break down the overly polished, uniform structure of AI-generated text.

Instead of: “The study shows little empirical follow-up, although it is widely cited.” Try: “Although widely cited, the study has received little empirical follow-up.”

Let’s Humanize AI Text: A Step-by-Step Example

[08:36.467]

Let’s take a paragraph generated by ChatGPT. The prompt was: “please write a 100-word paragraph, academically about importance of self-esteem for migrants when they communicate in English that is not their first language.”

Here is the original AI-generated text:

“Self-esteem plays a critical role in shaping the communicative experiences of migrants using English as a second language. High self-esteem fosters confidence, which is essential for engaging in conversations, expressing needs, and participating in social, educational, and professional contexts. Conversely, low self-esteem may hinder communication by increasing anxiety, fear of judgment, and reluctance to speak, ultimately affecting language development and social integration. Migrants who perceive their English skills as inadequate may internalise feelings of inferiority, reinforcing exclusion and marginalisation. Therefore, supporting self-esteem is vital not only for linguistic competence but also for promoting migrants’ overall well-being and inclusion in host societies.”

[08:57.962]

When run through an AI detector, this text is flagged as 100% likely to be AI-generated. Now, let’s manually edit it, sentence by sentence, applying the principles discussed earlier.

[09:03.957]

1. First Sentence: “Self-esteem plays a critical role…” is too absolute. We add hesitation: “Self-esteem can play a critical role…” We also add a citation, which is standard for academic writing and helps break the flow: “(Jackson, 2020).”

2. Second Sentence: The original is long and lists many things. We’ll simplify and rephrase. “High self-esteem fosters confidence…” becomes “It has also been shown that high self-esteem can increase confidence and as a result, their willingness to participate in various conversational contexts.” This tones down the claim and varies the sentence structure.

3. Third Sentence: This sentence is complex and lists multiple negative effects. We’ll simplify and rephrase to make it more direct and less like a list. “Conversely” is a typical AI transition word, so let’s change it to “Low self-esteem, on the other hand, may affect social integration, mainly due to increased anxiety and fear of judgement.” This focuses the sentence and removes some of the AI-like filler.

4. Fourth and Fifth Sentences: These two sentences are closely related. The original is wordy and predictable. We can combine and simplify them while adding nuance. Let’s rephrase it to focus on the key ideas: “As low self-perceived English skills are believed to lead to inferiority complex and exclusion, it is crucial to support self-esteem building to promote the migrants’ linguistic and social inclusion.” This combines ideas, changes sentence structure, and uses more academic phrasing like “are believed to lead to.”

After these changes, here is the revised paragraph:

“Self-esteem can play a critical role in shaping the communicative experiences of migrants using English as a second language (Jackson, 2020). It has also been shown that high self-esteem can increase confidence and as a result, their willingness to participate in various conversational contexts. Low self-esteem, on the other hand, may affect social integration, mainly due to increased anxiety and fear of judgment. As low self-perceived English skills are believed to lead to inferiority complex and exclusion, it is crucial to support self-esteem building to promote the migrants’ linguistic and social inclusion.”

[15:36.574]

When this revised text is run through the same AI detector, it scores 0% likely to be AI-generated. The key was not just changing words but fundamentally altering the structure, rhythm, and certainty of the sentences. By understanding what AI detectors look for, you can manually edit AI-generated text to be both undetectable and of higher academic quality.