Humanizer: Teaching AI to Stop Writing Like AI

AI writing assistants made everyone more productive. They also made everyone's writing sound the same. The prose carries telltale markers—formulaic "it's not X, it's Y" constructions, bolded bullet points, dramatic one-line closers—that readers now spot instantly. A Hacker News commenter captured the exhaustion: these patterns make otherwise useful writing difficult to read.

Humanizer tackles the problem by using an LLM to critique LLM output. The tool checks text against 26 recognizable AI-writing patterns—staged run-ups, inflated significance, decorative formatting, chatbot residue, repeated explanations—and rewrites to eliminate them. With nearly 60,000 GitHub stars, the repository suggests developers and writers share the same aesthetic fatigue.

How it works: Wikipedia as style guide

The architecture is straightforward: Humanizer bases its checks on Wikipedia's Signs of AI writing guide, then runs text through detection, drafting, critique, and final-review passes while preserving source claims and factual details. The irony is intentional—asking a language model to stop writing like a language model. The approach works because it addresses reader experience, not just detection evasion.

The Wikipedia foundation matters. Rather than inventing arbitrary style rules, the tool anchors to patterns that experienced editors already identified as problematic in real-world content moderation. That grounding gives the 26-pattern checklist legitimacy.

The detection controversy

Can prompt-based rewriting actually fool statistical detectors? In issue #82, Rephrasy argued no—output remains statistically predictable because it uses the same base model, so detectors examining perplexity and burstiness may continue to flag the text. The maintainer closed the issue as commercial detector-bypass promotion, but the technical question stands.

A Reddit tester reported mixed results: one sample passed ZeroGPT, another still looked like typical AI writing and failed detection testing. These inconsistencies reflect what Humanizer tackles—a tool working through a hard problem. The legitimate use case stands on its own: making AI-assisted writing more pleasant to read, regardless of whether it evades detectors.

Both AI writing tools and detection systems serve real purposes. The former boost productivity; the latter protect academic integrity. Humanizer occupies the space between, aiming to preserve the productivity gain while reducing reader fatigue.

Results vary, adoption grows

The technical debate hasn't slowed adoption. As of November 2025, the repository had accumulated 59,649 stars, 4,700 forks, and 25 contributors across six releases. The MIT license enabled rapid expansion—Hermes Agent bundled a ported creative-humanizer skill, and the repository forked into a separate project that preserves its portable SKILL.md installation model. Recent pull requests show proposed support for Pi, Google Antigravity, Gemini, and Russian-language use.

The inconsistent results reported by testers signal that Humanizer is still finding its footing—typical for a repository growing this fast. Rewriting quality varies, and the tool doesn't guarantee detection evasion. But 59,000 stars suggest it scratched a real itch.

Who this is for

Humanizer targets developers and writers using AI assistants like ChatGPT, Claude, or Cursor who want output to sound less robotic. If you've found yourself editing out the same stilted patterns from every AI draft, this tool automates part of that polish. It won't replace editorial judgment, but it saves time on the first pass.

The repository respects both sides—appreciating AI writing tools for their utility while acknowledging that their output needs refinement for human readers. That balance matters more than winning the detection arms race.