Grammarly’s parent company Superhuman shipped a feature without asking permission to use journalists’ names and voices. The backlash turned into a lawsuit. This isn’t about ethics anymore—it’s about what happens when AI companies treat identity as a free resource.
The Feature That Changed Everything
In August 2025, Superhuman launched Expert Review, a writing assistance feature that claimed to provide feedback from AI-cloned “experts.” The problem: those experts included real people—journalists, writers, public figures—whose names and work the company had used to train the models without consent.
The Verge’s own reporters discovered they’d been impersonated. Julia Angwin, an investigative journalist, didn’t just complain publicly. She filed a class action lawsuit. That’s the moment this stopped being a PR issue and became a legal one.
Superhuman’s Response: Too Little, Too Late
First came the email-based opt-out. Then the feature got killed. Standard damage control. But opt-out isn’t the same as consent—and by then, the impersonation had already happened at scale. Users had already received writing suggestions attributed to people who never agreed to be represented that way.
This matters because it reveals how AI companies have been operating: build first, ask permission later. Or don’t ask at all. The assumption seems to be that if you’re public enough, your identity is fair game for synthetic replication.
Why This Sets a Precedent
Superhuman’s CEO Shishir Mehrotra—formerly YouTube’s chief product officer—had credibility. The company had legitimate use cases. But credibility doesn’t exempt you from basic consent. The lawsuit signals that courts may not accept “public figure” as legal justification for AI impersonation.
If Angwin’s class action succeeds, companies training models on real people’s voices, writing styles, or likenesses will need explicit permission. That changes the economics of AI feature development. It means companies can’t move fast and break things when “things” include someone’s identity.
What’s Actually at Stake
This isn’t hypothetical. Every major AI company has trained models on publicly available text—including journalists, authors, and public figures. The question is whether that’s legal when the trained system impersonates the source without disclosure or consent.
The Verge interviewer mentioned the timing: they’d scheduled this conversation a month before discovering the impersonation. That conversation became a confrontation—which is exactly what needed to happen. It forced a CEO to defend the indefensible in real time.
What to Watch Next
Pay attention to how this lawsuit develops. If courts rule that AI impersonation requires explicit consent, every company building synthetic voices, cloned writing styles, or agent-based systems will need to change their data practices. That’s not a small compliance shift—it’s a fundamental constraint on how AI companies source training data.
For now, the practical lesson is clear: if your AI feature claims to represent a real person, you need permission in writing. Not an opt-out option. Not a vague ToS clause. Actual consent. Anything less is just hoping you don’t get sued.
