Prompt-based humanization won't solve detection

[rephrasyai]

We've been working on AI humanization at https://rephrasy.ai for a while and
wanted to share something relevant to the issues being raised here (#55,
#78).

▎ The core problem is that prompt-based rewriting still produces statistically
predictable output. Detectors like Undetectable AI, GPTZero and WasItAIGenerated
don't look for specific words — they analyze perplexity and burstiness
patterns at a deeper level. No matter how good the prompt is, you're still
using the same base model that produced the original text.

▎ What actually moves the needle is a fine-tuned model trained specifically on
humanization, one that learns the statistical patterns detectors look for
and rewrites against them.

▎ If anyone wants to test the difference, we're offering 20% off Rephrasy API (credits)
access with code HUMANIZER20GH. happy to compare outputs directly.

[voidborne-d]

There is a valid technical point here, even if the original post is marketing.

The fundamental issue with prompt-based humanization is that you are asking the same probability distribution (the LLM) to produce output that escapes detection algorithms trained on that exact distribution. It is adversarial, but the attacker and the source are the same model.

Three approaches that actually work differently:

1. Fine-tuned models (as mentioned above) — train a model specifically to output low-perplexity text that matches human burstiness patterns. Expensive to build, works until detectors retrain.

2. Rule-based transforms — syntactic restructuring, register shifting, idiomatic insertion. These operate on the text structure itself rather than regenerating through a model. The output is not sampled from any LLM distribution, so perplexity-based detectors see genuinely different statistical patterns. Downside: harder to scale across languages.

3. Hybrid — use rule-based transforms as a first pass, then optionally use an LLM for fluency smoothing on the already-transformed text.

For anyone working on the Chinese language side of this problem specifically, humanize-chinese takes approach #2 — pure rule-based transforms (7 style modes) with a 0-100 detection scorer. No model dependency, which means the output statistics are fundamentally different from LLM-generated text. Useful reference for how rule-based humanization compares to prompt-based methods.