When your writing tool and your AI detector are the same company

On June 24, 2026, the company that owns Grammarly bought GPTZero. Superhuman (Grammarly's parent since its October 2025 rebrand, with 40 million daily users) acquired one of the best known AI detectors and announced it will be folded into Superhuman Go, the assistant that already runs across roughly a million apps and websites. They are calling it an "authenticity layer."

The company that helps you write is now also the company that decides whether your writing looks like AI. Those two jobs used to belong to separate companies. As of today, for a lot of people who write for school or work, they belong to the same one.

Why one company doing both is a problem

A writing assistant is built to help put words on the page. A detector is built to flag writing that looks machine-made. One company now runs both, with the detector switched on by default wherever you write. So the same company that helped write your sentence also decides whether it looks like AI.

When a detector flags a sentence you actually wrote yourself, who in that loop is on your side?

More surfaces, more false positives

AI detectors get genuine human writing wrong about 1 in 20 times. Even the strong ones, held to a strict limit on false positives, miss that mark sometimes. That rate has not gone anywhere. What changed today is the surface area.

Detection used to be something you opted into. You pasted text into a box when you were worried. Spreading it across a million apps makes it ambient: it runs whether or not you asked it to. More writing scanned, more often, by default, means more honest sentences flagged in raw numbers, even if the error rate holds perfectly steady. And the people it hits hardest are the same as always: students, non-native English speakers, and anyone who writes in a clean, formal style.

Where GPTypo fits, and where it doesn't

GPTypo runs a detection model of its own, and it suggests rewrites, so we're not pretending to stand outside this business. What's different is who we answer to, and what the tool is for:

• We don't own both the tool grading you and the tool you write in. GPTypo isn't bundled into your inbox or your editor, scanning by default. You bring it your text when you want a read, and that is the only time it does anything.
• The product exists to catch the wrong flag. We earn our keep when someone who wrote the thing themselves can see what a detector is reacting to and adjust it in their own voice. We don't sell the assistant that creates the suspicion in the first place, and we have no interest in a one-click trick for faking your way past detection.
• We publish where we land. Our model sits on the public RAID leaderboard from ACL 2024, scored on a hidden test set we had no hand in grading. So does GPTZero's. So, for the record, does Superhuman's, and their detector scores well. We're glad to be measured on a test none of us set. Here is how that benchmark works and why a public result beats a vendor's own number.

Independence used to be a nice-to-have for GPTypo. Today it became the whole point. When the writing tool and the detector answer to the same company, a writer needs at least one tool in the loop whose only job is to be on their side.

A detector you don't control, built into a tool you didn't choose, isn't a second opinion. It's the first one, everywhere.

What this means for how you work

If detection is going to be ambient, treat it that way:

• Get a second opinion you control. Detectors disagree with each other constantly. If one tool flags you, run the same passage through a different one before you accept the verdict.
• Keep your own receipts. Version history, earlier drafts, and notes are timestamped proof of how a document actually came together. The useful version of provenance is the one the writer holds, not the one stored inside the company doing the grading. Here is the full playbook for handling an accusation.
• Check before you submit, not after. Seeing what a detector reacts to while you can still edit beats arguing about it once someone else has run the scan.

GPTypo is one way to get that second opinion. Paste a draft in and click Verify for a scan from a model we benchmark in the open, with the exact sentences and reasons behind the score, so you decide what to change and what to keep. The flag is information. It was never meant to be the verdict.