Open source AI humanizers: three GitHub projects and what each does

Two kinds of project carry the label. A rule-based skill is a text file of instructions that a model you already use reads before rewriting; it has no model of its own and no server. A pipeline is code that calls models and services with your keys and returns the rewritten text; it can run on your machine or a server. Both perform the same operation as a hosted humanizer, a model rewriting text, and both expose the steps.

The research behind the second kind is public too. Krishna and colleagues built DIPPER, an 11 billion parameter paraphraser that reorders content and controls lexical diversity, and reported that it dropped DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). The paper measured detectors; it judged meaning by automatic measures, not by editors, which is the gap every project below also leaves open.

blader/humanizer is the most installed of the three and the simplest: one SKILL.md that an agent reads when asked to humanize text. The README describes 26 patterns, grouped in six sections and numbered by strength, drawn from Wikipedia's "Signs of AI writing" page: staging instead of stating, rhythm by rule, inflation and borrowed authority, formatting by rule, leftovers from the chat, and writing for the wrong reader (blader/humanizer README, read 9 November 2026). The skill marks every tell, drafts a rewrite, checks the draft against the patterns and the original claims, and writes a final version.

Two sentences in the README set it apart from the hosted tools. It does not make things up: a name, number, date or quote must come from the source or the writer, and the skill asks when a detail is missing. And it edits for human readers; the README states that getting past AI detectors is not a goal and that detectors still flag most of its output. Installation is two plugin commands in Claude Code, a ZIP upload in Claude.ai, or one npx command for other agents. The full description, pattern list and install routes are in the Claude humanizer skill guide.

lynote-ai/humanize-text is a pipeline, not a prompt. Its README describes four steps: a high-temperature model rewrite from the input into Chinese, a second model rewrite from Chinese into Japanese that carries the first step as history, a machine translation hop from Japanese into Finnish, and a second hop by a different translation engine from Finnish back into English (lynote-ai/humanize-text README, read 9 November 2026). The README's stated reason is that two rewrites at high temperature break the statistical patterns of AI text, and two translation engines across distant languages add structural change that no single engine's fingerprint survives.

The project runs three ways: a Python command line, an n8n workflow file for no-code use, and a Docker compose file that exposes a REST endpoint. It needs Python 3.10 or newer, a configuration file, a model key (DeepSeek by default, with OpenRouter and other OpenAI-compatible endpoints listed) and a Niutrans key. Three tiers are named: standard (the four steps), advanced (added rounds of model rewriting) and focus (a detection-guided feedback loop). The README's own before-and-after example shows a two-sentence quantum computing paragraph whose wording changes while the claims hold, and the same repository links a hosted version run by the project's company.

Two cautions follow from the design. Four translation hops through three languages is the most meaning-moving operation on this page, and the README does not publish an editor-rated meaning figure. And the per-run cost is your own: model tokens plus translation characters, with no monthly cap because there is no vendor pool.

harshaneel/humanize bundles two rule-based skills, humanize and ai-check, with no models trained, no API calls and no detector in the loop; the README calls the skill "a rulebook the host LLM follows" (harshaneel/humanize README, read 9 November 2026). The humanize skill applies nine levers the README says are drawn from the detection literature; the ai-check skill scores nine signal categories and returns a verdict with a confidence. It installs through a Claude Code or Codex plugin marketplace, or an install script that places the skills in the agents' skill folders.

Its most useful sentence is a limit. The README states that a rule-based approach is competitive against open-source perplexity-based detectors but not against learned commercial classifiers such as Pangram or GPTZero, and it keeps a section on what the approach cannot do. That is a narrower claim than any hosted vendor makes about itself, and it matches the published record: detector makers now train against paraphrasing, and a 2023 reading says little about a 2026 model.

The published method scores every tool on meaning kept (35 of 100), errors introduced (20), detector pass rate (30) and limits and access (15). Open source projects win the last field outright: no request cap beyond the host's context window, no monthly pool, no sign-up, no price. On the other three they have the same gap as the hosted tools, no editor-rated meaning figure, with one difference: two of the three say in writing that detector readings are not their promise.

For a developer who wants the operation inside their own product, the pipeline's REST endpoint and the vendor APIs are compared in the humanizer API guide. For a writer who wants a free allowance with no setup, the standing free tiers are on the free humanizer list. No detector score should be the only evidence in a misconduct case, and students should check their institution's AI policy before using any of these; open source does not make undisclosed AI use acceptable. The rest of this series is in the guides hub.