'AI Detector and Fixer' Tools: What They Really Do

In June 2026, Superhuman, the parent company of Grammarly, acquired the AI detector GPTZero, which had grown to 19 million registered users and roughly $30 million in annual recurring revenue, per TechCrunch. Detection is a real industry now. And trailing right behind it is a stranger product category: the "detector and fixer" bundle, which sells the alarm and the lockpick in the same box.

An AI detector and fixer is a bundled tool that first scores your text for AI probability, then rewrites it until the score drops. The detection half is standard technology. The "fixing" half is a paraphraser built to evade detectors, and it usually trades away quality, accuracy and integrity to buy that lower number.

What an AI detector and fixer actually sells

The bundle has two parts. The detector half is a classifier that estimates how machine-like your text looks, the same statistical machinery every serious tool uses; our guide to how AI detectors work walks through it. Nothing exotic there.

The fixer half is where the pitch lives. It's a paraphrasing model instructed to rewrite text so that detectors score it as human: swap word choices, restructure sentences, roughen the rhythm, sometimes sprinkle in small irregularities that classifiers read as human noise. Marketing copy calls this "humanizing." Mechanically, it is adversarial rewriting, and the target is a number rather than a reader.

Notice the business model. The detector manufactures the anxiety, the fixer sells the relief, and both charges land on the same subscription. It's the mosquito and the repellent from a single vendor.

The loop: scan, rewrite, rescan

Here's what actually happens when you paste an essay into one of these tools. The built-in detector scores it, say 85% AI. The rewriter produces a new version. The same detector rescans, and the loop repeats until the in-house number falls under some threshold, at which point you're shown a reassuring green badge.

The catch is the phrase "the same detector." The fixer is optimized against its own sibling, so what you've proven is that vendor A's paraphraser can fool vendor A's classifier. Turnitin, GPTZero, Copyleaks and everyone else train on different data with different thresholds, and whether a rewrite that fools one fools the others is an open question every single time. Our long answer on whether ChatGPT output stays detectable covers why detectors disagree so much in the first place.

Picture the transaction from the buyer's side. A student pastes a chatbot-drafted essay, watches 85% fall to 6%, and pays for the download believing the number travels with the file. It doesn't. The 6% describes one classifier's reaction to one rewrite on one afternoon. The essay then meets a different detector at school, trained on different data, and whatever happens next was never part of the sales demo. The badge was real; the promise attached to it was not.

Does the fixing actually fool detectors?

Honest answer: more often than detector vendors like to admit, and less reliably than fixer vendors claim.

The strongest public evidence is the RAID benchmark (Dugan et al., ACL 2024), which tested detectors against more than 10 million documents and 12 adversarial attacks. Commercial detectors degraded sharply under paraphrase and homoglyph attacks. That's the fixer's case in one sentence: rewriting works often enough to be a genuine problem for detection.

Now the other side. A 2025 ACL study by Russell, Karpinska and Iyyer found that expert annotators who frequently use ChatGPT identified AI-generated text with 99.3% accuracy, and stayed accurate against evasion tactics that beat the automated tools. Machines can be gamed with synonym math. An experienced reader noticing that your essay suddenly has no voice, no specifics and three garbled idioms cannot. We keep a running evidence file in can AI text really be made undetectable, and the pattern holds: evasion beats software far more easily than it beats people.

What the rewrite costs you

Fixer output has a signature, and it isn't "human." It's mush. Paraphrasers swap precise terms for approximate ones, flatten whatever voice the draft had, and garble numbers, citations and idioms, because the rewriting objective is statistical, not semantic.

The marketing promise	What typically happens
"Bypass any detector"	Beats the bundled one; others vary scan to scan
"Preserves your meaning"	Precise terms and citations come out warped
"Reads naturally"	Experienced readers still notice; 99.3% in one ACL study
"One click"	In practice, a loop of rewrites and rescans
"Safe to submit"	Masked AI text is still AI text under most policies

The integrity cost is simpler. Academic and workplace policies judge where the work came from, not what a scanner said about it. Running prohibited AI text through a fixer doesn't launder the origin; it adds a deliberate concealment step, which is the part misconduct panels treat most harshly.

There's also a quieter irony: chasing 0% is chasing a number honest writing doesn't reliably produce either. Human text gets flagged too, for reasons we unpack in why human writing gets flagged, so the target these tools sell is fake precision in both directions.

Why we don't sell a humanizer

AI Detector 360 gets asked for one constantly, and the answer stays no, for a reason worth spelling out: a company selling both detection and evasion has no incentive for either product to work. Every fixer sale depends on the detector being scary, and every detector improvement breaks the fixer you sold last month. One of those products is always quietly betraying its customers.

So we do detection only, and we do it with the error bars showing: multi-engine scoring instead of one model's opinion, a sentence-level heatmap so you can see what drove the score, and an explicit confidence label on every scan. How those are calibrated is public on our methodology page. No detector is proof, ours included; a score is evidence to weigh, and we'd rather tell you that plainly than sell you an antidote to our own product.

Legitimate ways to use the detector half

There is a version of "detect and fix" that's completely fine: the version where you do the fixing.

Say you drafted a report yourself but leaned on formulaic phrasing, or you used AI for an outline in a course that permits it and wrote the prose on top. Scanning before you submit tells you how the text reads statistically. If sections light up, rewrite them the honest way, then move on with your life instead of iterating toward a magic zero.

What the honest rewrite looks like, concretely:

• Replace boilerplate claims with specifics only you know: the dataset's actual size, the source's actual name, the thing that surprised you.
• Break the metronome. Combine two short sentences, split a long one, let a fragment stand.
• Cut the throat-clearing openers and summary sentences a chatbot loves; start where the point starts.
• Read one paragraph aloud. Anywhere your voice goes flat, the writing probably did too.

Notice none of that requires software, and all of it makes the writing better, which paraphrase loops reliably don't. A detector score falling as a side effect of genuine revision is the system working. A score falling while the text gets worse is the tell that you're laundering, not editing.

That pre-flight check is what our free AI detector is for: 5,000 characters per scan, no sign-up, heatmap and confidence level included. The line between revising and laundering isn't subtle. It's just the answer to one question: who did the writing, you or the loop?