Ghost-2: Inside Our New AI Humanization Model

Every humanizer marketing page shows you a pass rate. Almost none of them tell you when it was measured, and that date matters more than the number.

AI detectors are not static. GPTZero publicly documents that it trains specifically against humanized text, and Originality.ai ships new model versions on a regular cadence. Every one of those updates is aimed at exactly one thing: catching text that used to pass. So a humanizer is not a product you evaluate once. It's a product in a permanent arms race, and its scores have a half-life.

We measured that half-life on ourselves. Re-running Ghost-1 through the same evaluation months after launch, its mean pass rate had dropped from 96.2% to 93.1%. GPTZero was the sharpest fall (92.5 to 86.2), Originality.ai and Quillbot both clawed back several points, and only ZeroGPT and Grammarly stood still. Ghost-1 was still strong, and for most everyday use those numbers held up fine. But the trend line only points one direction, and for our users the stakes of a single flag are not abstract: a flagged essay, a rejected article, a client asking uncomfortable questions.

There are two ways to respond to decay. Most tools patch: tweak prompts, chase whichever detector update hurt them last month. We took the slower route and retrained the model itself on what detectors have become, not what they were in 2025.

Ghost-2 is a custom-trained text-to-text humanization model, the second generation of the Ghost family. It takes AI-generated text and rewrites it to read the way a person actually writes: uneven sentence rhythms, deliberate word choices, the small structural irregularities that human writing has and language-model output flattens away.

The design philosophy hasn't changed since Ghost-1, because it's the reason the approach works at all. Most humanizers are wrappers: a general-purpose model, a paraphrasing prompt, and aggressive sampling settings that inject randomness until detectors get confused. That trick buys detector evasion by making text weirder, which is why so much humanized text reads like it was translated through three languages. Ghost models are trained to produce text that is statistically human in the first place, so passing detectors is a side effect of the writing being genuinely better, not a trick played on a specific detector's current weights.

What did change in Ghost-2 is the training bar. Ghost-1 was tuned overwhelmingly for evasion, and it showed in the output: it passed detectors convincingly while sometimes reading rougher than we wanted. Ghost-2 was trained against both objectives at once, with modern detector behavior in the loop and writing quality scored on every output. The result is the first Ghost model where the quality numbers moved as much as the detection numbers.

The numbers in the next section come from the same pipeline we've used since Ghost-1, so generations are directly comparable. Five steps:

1. A large fixed sample of AI-generated English texts spanning essays, technical documentation, blog posts, and conversational prose.
2. Each text is rewritten once by the model on default settings. No per-detector tuning, no retries, no best-of-N cherry-picking.
3. Outputs are scored independently by six detectors: GPTZero, Originality.ai, ZeroGPT, Turnitin, Quillbot, and Grammarly. We never see detector scores during humanization.
4. Outputs are also scored by general-purpose LLMs on whether the rewrite preserves meaning, reads fluently, and stays consistent across the document.
5. A text passes a detector if it returns under 30% AI probability. The mean pass rate averages across detectors, not texts.

Two honest caveats. The Turnitin column runs on a smaller sample than the others, since Turnitin has no public API and every check is manual. And these are our own numbers on our own eval set; that's exactly why the independent benchmark section further down exists.

Here's the full table. The interesting comparison isn't Ghost-2 against Ghost-1's launch numbers, it's Ghost-2 against Ghost-1 as it performs today, after detectors have had months to adapt. That's the column that answers the question a returning user actually has: what do I gain by switching right now?

A fair question at this point: for how long? Detectors will keep updating, and these numbers will drift the same way Ghost-1's did. The difference is that we now publish the drift (the Ghost-1 "today" column above is exactly that), and the evaluation pipeline reruns continuously, so the next retrain starts from live data rather than a stale snapshot.

Pass rates get the headlines, but a rewrite that passes every detector and mangles your argument is worthless. Research backs up how real this tension is: TH-Bench, an academic benchmark that tested humanization attacks against thirteen detectors, found that no method scored well on evasion and text quality at the same time. That tradeoff is the category's central problem, and it's where Ghost-1 left the most room to improve.

We score quality with general-purpose LLMs judging three things on the shared eval set: whether the rewrite preserves the original meaning, how naturally it reads, and whether tone and quality stay consistent across a full document. Here's Ghost-1 against Ghost-2:

Self-reported evals from the company that built the model deserve skepticism, ours included. So here's the external check.

AI Humanizer Benchmark runs a public leaderboard that scores humanizers against seven detectors (GPTZero, Originality.ai, Winston AI, Copyleaks, ZeroGPT, QuillBot, Grammarly) on a standardized test set, 33 samples per tool, with a published methodology. In the September 2026 cycle, UndetectedGPT running Ghost-2 ranks #1 of 11 humanizers, with an overall score of 84.9. The component scores behind it: 86.3 on detector bypass, 84.8 on meaning preservation, 77.3 on readability.

The interesting part isn't the rank, it's the shape of the scoreboard underneath it. The highest raw bypass score in the entire field belongs to WriteHuman at 90.4, and it finished third overall, because its meaning-preservation score collapsed to 76.2. That's the stealth-versus-quality tradeoff we described earlier, showing up on someone else's test set with someone else's methodology. Most tools in the field bought their bypass score by letting the writing degrade. Ghost-2 took the top spot the other way: near the top on stealth while holding the strongest quality-side scores of any high-bypass tool in the cycle. The "both axes at once" claim from our own eval is exactly what the leaderboard shows.

Sharp-eyed readers will notice the benchmark's bypass score (86.3) sits well below the 96 to 99% pass rates in our table above, so let's address that directly rather than hoping nobody asks. The test sets are different by design. The benchmark deliberately mixes in hard-to-humanize inputs to separate the field; its own writeup notes that short, formulaic text is far harder to humanize than long-form prose, and every tool's scores run lower there than in everyday use. Our eval set is built to reflect typical real-world text: essays, articles, documentation. Both numbers are true; they're answering different questions. What matters for a ranking is relative position on identical inputs, and on identical inputs Ghost-2 comes out first.

If you'd rather run your own check, that's the best benchmark of all: run your text through the humanizer, then paste the output into whichever detector you're worried about. The free tier exists precisely so you can do this before paying us anything.

The arms race doesn't pause because we shipped. Detectors will adapt to Ghost-2 the way they adapted to Ghost-1, our continuous evaluation will catch the drift as it happens, and the next Ghost model is already in the pipeline with the same two-axis bar: better writing and better evasion, together or not at all.

The commitment we can make is the one this post models: current numbers, published decay, independent verification, and a family page that tracks every generation honestly. The lineage from Ghost 0.7 through Ghost-2 is documented there, launch scores and today-scores side by side. If a future number goes down, you'll see it there before a competitor tells you about it.

For the wider context on how Ghost-2 compares to other tools on the market, our best AI humanizers guide covers the field.