Bypass GPTZero with HumanGPT.

GPTZero's detection model relies on two core metrics that separate human writing from machine output.

The first is perplexity. In simple terms, perplexity measures how predictable the next word in a sentence is. When a language model generates text, it picks the most statistically likely word at each step. That makes the output low-perplexity: predictable, smooth, and boring to a detection algorithm. Human writers are different. We reach for unusual word choices, use idioms that don't make literal sense, start sentences with conjunctions for no grammatical reason. That unpredictability registers as high perplexity.

The second metric is burstiness. This measures variance in sentence complexity across a passage. AI models produce sentences that are roughly the same length and structure. Paragraph after paragraph of 15 to 20 word sentences, each with a subject-verb-object core and one dependent clause. Humans don't write like that. We write a two-word fragment. Then a 35-word sentence that wanders through three parenthetical asides before landing on the verb. That variance is burstiness, and GPTZero is specifically looking for its absence.

GPTZero also runs a classifier trained on millions of human and AI text samples. This classifier picks up on subtler patterns: the way AI models handle transitions between paragraphs (they love "moreover" and "furthermore"), the absence of contractions in formal output, and the tendency to open every paragraph with a topic sentence followed by three supporting sentences in identical structure.

The result is a confidence score from 0 to 100. Anything above roughly 85 gets flagged as "likely AI generated." The gray zone between 50 and 85 gets marked as "mixed." And below 50 passes as human. HumanGPT's job is to push your output solidly below that threshold.

We test every Monday with fresh samples across all three modes (Light, Medium, Heavy). Here's what we consistently see with a 250-word academic paragraph generated by GPT-5.5.

Before humanizing, a typical ChatGPT output scores 92 to 97 on GPTZero's AI probability scale. That's deep in the "AI generated" zone. The perplexity is low (around 15 to 25, where human writing typically sits at 60+). Burstiness is flat (variance under 5, where human text shows variance of 15 to 30+).

After running through HumanGPT on Medium mode, the same content scores between 8 and 22 on GPTZero. Perplexity jumps to 55 to 80. Burstiness variance climbs to 18 to 35. The text reads as something a university student would actually write: competent, clear, slightly informal in places, with natural rhythm.

Heavy mode pushes it further. Scores typically land between 2 and 12, with perplexity above 70 and burstiness well into the human range. At this point, GPTZero classifies it as "human" with high confidence.

Light mode is more conservative. It keeps about 70% of the original wording and focuses on the most obvious AI tells. Scores typically drop from 95 to somewhere between 30 and 50. That's borderline. For high-stakes submissions (final papers, published articles), we recommend Medium or Heavy.