Ghost-2
Our new humanization model that rewrites AI text to read like it was written by a human.
Ghost-2 at a glance.
The numbers behind the model. What it is, what it does, and where it runs.
- Model family
- Ghost
- Version
- 2.0
- Type
- Humanization (text-to-text)
- Languages
- EN primary · 100+ supported
- Detectors covered
- 10+ (GPTZero, Originality.ai, ZeroGPT, Turnitin, Quillbot, Grammarly, and more)
- Mean pass rate
- 98.3%
- Released
- Live
- Previous version
- Ghost-1 · available via API and model settings
Tested where it counts.
Every Ghost release runs the same eval: pass rate (scored <30% AI) on a large fixed set of unique English texts, scored independently by six detectors.
at launch (%)Ghost-1
today (%)Ghost-2 (%)Δ
Evaluated on deduplicated English-only texts, each scored by six independent detectors. "Pass" = detector returns <30% AI probability. *Turnitin was evaluated on a smaller sample due to its gated access. Ghost-1 launch figures come from its original eval run; several detectors have since updated, which lowered Ghost-1's rates today. Δ compares Ghost-2 with Ghost-1 today, on the same test set.
Output quality
Stealth is only half the eval. Every release is also judged on how the writing holds up, and Ghost-2 improved on all three fronts.
Quality is scored by general-purpose LLMs on the same eval set: each rewrite is judged on preserved meaning, fluency, and stability across runs.
of 11 humanizers · September 2026
GPTZero · Originality.ai · Winston AI · ZeroGPT · Grammarly · QuillBot · Copyleaks
Meaning · readability · consistency
Ranked #1 of 11 humanizers in the September 2026 AI Humanizer Benchmark, scored across 7 detectors with a fixed public methodology.
See the leaderboardWhy are the pass rates above higher than the benchmark's scores?
The test sets are different. The benchmark scores every tool on deliberately hard-to-humanize inputs, so scores there run lower for everyone. The pass rates above are measured on text typical of everyday use.
From Ghost-1 to Ghost-2.
Detectors change constantly, so a model frozen in time slowly goes stale. Every Ghost release re-runs the same benchmark, and the next model is already in the works while the current one ships.
- nextIn training
The next Ghost
Training never stops. The next generation is already in the works against the newest detector releases.
- v2.0Now
Ghost-2
Second-generation model, retrained end to end against the newest detector versions. Mean pass rate 98.3% across six detectors, up from Ghost-1's 96.2%. Writing quality moved with it: tighter meaning preservation and steadier output across writing styles.
- v1.0Previous generation
Ghost-1
First stable release. Mean pass rate 96.2% across six detectors. Still available through the API and model settings.
- v0.9Last cycle
Tone preserve
Reduced semantic drift on long-form academic and technical text.
- v0.8Earlier
Turnitin push
Added Turnitin to the eval set; pass rate moved 91 → 97%.
- v0.7Earlier
Citation lock
References, footnotes, and inline citations are no longer rewritten.
How we evaluate.
The same eval runs every release. No per-detector tuning, no hand-picked samples.
- 01
Source
A large sample of AI-generated English texts spanning essays, technical docs, blog posts, and conversational prose.
- 02
Humanize
Each text is rewritten by Ghost-2 with default settings (no per-detector tuning or retries).
- 03
Detect
Outputs are scored independently by six detectors. We never see scores during humanization.
- 04
Score quality
Outputs are also scored by general-purpose LLMs to check that each rewrite preserves meaning, structure, and fluency.
- 05
Pass / fail
A text passes a detector if it returns less than 30% AI probability. Mean pass rate is averaged across detectors, not texts.
Two generations in, still one job:sounding human.
Ghost-2 is the second generation of Ghost, our proprietary humanization model. Like every Ghost model, it rewrites AI-generated text so it reads like it was written by a human, bypassing every major AI detector while preserving meaning, tone, and fluency.
It is now the default model behind every humanization on UndetectedGPT: the web app, the API, and Ghost Writer. Ghost-1 stays available for anyone who wants to keep their existing setup.
Built to beat detectors
Ghost-2 isn't a generic paraphraser. It was purpose-built to bypass GPTZero, Originality.ai, ZeroGPT, Turnitin, Quillbot, Grammarly, and more.
Preserves meaning and tone
Ghost-2 rewrites how something is said, not what is said, keeping your original meaning, structure, references, and citations.
Continuously improving
As detectors evolve, so does Ghost-2. We constantly evaluate and update the model to stay ahead.
Live everywhere we ship.
Ghost-2 is the default for every UndetectedGPT user, in every surface. Ghost-1 stays available through the API and the humanizer's model settings.
Frequently asked questions
Ghost-2 is the second generation of UndetectedGPT's proprietary humanization model. It rewrites AI-generated text so it bypasses AI detectors like GPTZero, Originality.ai, ZeroGPT, Turnitin, Quillbot, and Grammarly while keeping your original meaning, references, and citations intact.
Yes. Ghost-2 is now the default everywhere, but Ghost-1 stays available: API users can pin it with the model parameter, and subscribers can switch models in the humanizer settings. Previous generations stay online so a new release never breaks your workflow.
Ghost-2 is benchmarked against the strictest AI detectors on the market: GPTZero, Originality.ai, ZeroGPT, Turnitin, Quillbot, and Grammarly. It runs the exact same eval as Ghost-1 and passes with a 98.3% average pass rate, up from 96.2%. Other detectors tend to be similar or easier to bypass.
No. Ghost-2 rewrites how something is said, not what is said. Your original meaning, structure, references, and citations are preserved.
The test sets are different. The benchmark scores every tool on deliberately hard-to-humanize inputs, so scores there run lower for everyone. The pass rates on this page are measured on text typical of everyday use.