Grammarly's Expert Review Debacle Exposes AI's Identity Crisis

Grammarly, now operating under the Superhuman brand after a corporate rebrand in October 2025, shipped a feature called Expert Review that let paying subscribers receive writing feedback ostensibly from Stephen King, Neil deGrasse Tyson, Kara Swisher, and hundreds of other real people. None of them agreed to participate. Some of them were dead. On March 11, 2026, facing a $5 million class action lawsuit filed by journalist Julia Angwin and a tidal wave of public outrage, the company pulled the feature. This is not just a product management failure. It is the first major test case for a legal and ethical boundary that every AI company will eventually slam into: who owns a person's expertise, voice, and intellectual identity in the age of generative models?

From Grammar Checker to Identity Laundering Machine

To understand how Grammarly ended up here, you need to trace the company's increasingly desperate pivot from a grammar tool to an AI productivity platform. In 2023, Grammarly introduced GrammarlyGO, its first generative AI feature, positioning itself against a rising tide of competitors: Microsoft Copilot embedded directly in Word, Google's Gemini woven into Docs, and a swarm of startups like Jasper, Copy.ai, and Writesonic eating the content generation market from below.

By mid-2024, Grammarly's core value proposition was eroding fast. Spell check and grammar correction had become commoditized features baked into every operating system and browser. The company needed to move upmarket. In July 2025, it acquired Superhuman, the premium email client, for an undisclosed sum. Then came the Coda acquisition. By October, Grammarly had rebranded the entire company as Superhuman Platform Inc., signaling an ambitious bet on becoming the connective tissue for AI-powered professional communication.

Expert Review was supposed to be the crown jewel of this new identity. Instead of generic AI writing suggestions, users could select a specific real-world expert and receive feedback styled after that person's known perspectives and editorial sensibilities. The pitch was seductive: imagine getting your draft manuscript reviewed by Stephen King, or your research paper critiqued by a leading academic in your field. The problem, of course, is that none of this was real. The feature was an LLM generating plausible-sounding advice while wearing the skin of a real human being.

What makes this particularly egregious is that Grammarly did not stumble into this by accident. The company curated a specific list of experts. It built dropdown menus with their names. It designed user interfaces that presented AI-generated feedback as if it carried the imprimatur of these individuals. This was not a hallucination problem. It was a product decision, made deliberately, to monetize the identities of real people as a subscription feature.

The Legal Frontier: Personality Rights Meet Generative AI

The Angwin lawsuit, filed in the U.S. District Court for the Southern District of New York, alleges violations of publicity rights, a legal doctrine that gives individuals control over the commercial use of their name, likeness, and persona. Historically, right-of-publicity claims have centered on celebrity endorsements, unauthorized merchandise, and lookalike advertising. But generative AI has blown the boundaries of this doctrine wide open.

Consider what Grammarly actually built. It did not use recordings of these experts' voices. It did not reproduce their copyrighted text. It did not even claim they had endorsed the product. Instead, it created synthetic behavioral proxies: AI systems that simulate how a person might think, write, and critique. This sits in a legal gray zone that existing publicity rights frameworks were never designed to address. The question the court will have to answer is whether simulating someone's intellectual persona for commercial gain constitutes a violation of their identity rights, even when no verbatim content is reproduced.

This case arrives at a moment when courts worldwide are scrambling to define the boundaries of AI and identity. In India, the Delhi and Bombay High Courts issued landmark rulings in 2025 protecting Aishwarya Rai Bachchan and Asha Bhosle from AI-generated deepfakes and voice clones. The European Parliament's Digital Creativity Integrity Act codified vocal likeness as a monetizable creative asset. In the United States, multiple states have expanded their right-of-publicity statutes, and the bipartisan push for federal AI transparency legislation has accelerated throughout early 2026.

The Grammarly case is different from, and potentially more consequential than, the voice cloning and deepfake cases that preceded it. Those cases involved direct reproduction of sensory attributes: a voice, a face, a physical likeness. Grammarly's Expert Review reproduced something more abstract: a person's intellectual style, their editorial judgment, their professional reputation. If the court rules that this kind of simulation is actionable, it would establish a precedent that extends personality rights into the domain of cognitive identity. That would reshape how every AI company thinks about personalization features.

The Competitive Fallout: Who Wins and Who Scrambles

The immediate winner here is Microsoft. Copilot's integration into the Office suite already makes it the default AI writing assistant for enterprise customers, and Grammarly's self-inflicted credibility wound removes a competitor from serious consideration in procurement conversations happening right now. Google benefits similarly, though its Gemini writing features remain less mature than Copilot's tight integration with Word and Outlook.

The losers extend well beyond Grammarly. Every AI company that has built or is planning persona-based features needs to reassess immediately. OpenAI's GPT Store includes custom GPTs that explicitly mimic the communication styles of public figures. Character.AI's entire business model is built on conversational simulations of real and fictional personalities. Inflection AI's Pi and similar personal AI assistants are designed to develop distinctive personas over time. If the Angwin lawsuit succeeds, each of these products faces potential exposure.

The startup ecosystem for AI writing tools also takes collateral damage. Jasper, Copy.ai, Writer, and similar platforms have competed partly on the promise of style matching and tone calibration. While most of these tools do not explicitly name real people, the underlying capability is the same: training on or prompting with patterns derived from identifiable individuals' work. A broad ruling on AI personality rights could force these companies to prove the provenance of every stylistic capability they offer.

The most interesting competitive dynamic, though, is what this does to the market for expert knowledge platforms. Companies like Masterclass, Cameo, and even Substack have built businesses around monetizing individual expertise with the explicit consent and participation of those individuals. Grammarly essentially tried to replicate this value proposition without the consent part. If the legal system enforces a consent requirement for AI-based expertise simulation, it creates a new market for licensed personality APIs: authorized digital twins that experts control and profit from. This is not hypothetical. Several startups, including Delphi AI and Cojourney, have been building exactly this infrastructure since 2024.

What Every AI Builder Should Learn From This

The tactical lesson is straightforward: do not ship features that use real people's identities without their explicit, documented consent. But the strategic lesson runs deeper.

Grammarly's Expert Review exposed a fundamental tension in how AI companies think about personalization. The temptation is enormous. LLMs can simulate anyone's style with reasonable fidelity given enough training data. The capability exists in every foundation model. The difference between a responsible product and a lawsuit is whether you choose to package that capability as a named-person feature or as a generic stylistic tool.

For founders building AI products, the framework should be:

• Never associate generated output with a real person's identity unless that person has opted in. This applies to names, likenesses, professional titles, and institutional affiliations. Generic style descriptors like 'academic tone' or 'conversational journalism' are fine. 'Feedback from Kara Swisher' is not.
• Build consent infrastructure before you build persona features. If your product roadmap includes any form of expert simulation, the consent and licensing pipeline needs to ship before the feature does. Opt-out is not consent. Grammarly's initial approach of requiring experts to manually request removal was indefensible.
• Treat identity simulation as a category distinct from content generation. Your legal review process for a feature that generates text in a generic voice is fundamentally different from one that simulates a specific person's judgment. The risk profiles are not comparable.
• Watch the regulatory environment, especially in the EU and at the U.S. state level. Tennessee's ELVIS Act, passed in 2024, was the first U.S. law explicitly protecting against AI voice cloning. Similar legislation targeting broader identity simulation is advancing in California, New York, and Illinois. Federal legislation is likely by 2027.

For engineers specifically, this incident highlights the importance of treating identity attribution as a first-class concern in system design. If your product's prompt templates include real people's names, those names should be sourced from a consent database, not hardcoded into a dropdown menu. Every AI system that references real individuals should have an auditable trail showing authorization.

The Prediction: Licensed Digital Twins Become a Billion-Dollar Market

Here is where this story goes next. The Grammarly lawsuit will likely settle. Superhuman Platform Inc. has deep enough pockets and enough reputational motivation to make this go away before a precedent-setting ruling. But the damage is done. The industry now knows that unnamed-person AI features are a litigation magnet, and the market will respond accordingly.

Within 18 months, expect to see a new category of infrastructure companies emerge: identity licensing platforms that serve as intermediaries between public figures and AI companies. These platforms will handle consent management, revenue sharing, quality control over how a person's digital twin behaves, and ongoing monitoring for unauthorized use. Think of it as ASCAP or BMI for cognitive identity, collecting royalties every time an AI system simulates someone's expertise.

The bigger shift is philosophical. For the past three years, the AI industry has operated on an implicit assumption that publicly available information about a person, their writing style, their known opinions, their professional identity, is fair game for model training and feature design. The Grammarly debacle marks the moment that assumption started dying. The next phase of AI product development will be defined not just by what models can do, but by what companies are legally and ethically permitted to do with the identities those models can simulate.

Grammarly built a feature that treated real people as interchangeable costume pieces for a language model to wear. The backlash was swift, the lawsuit was inevitable, and the lesson should be permanent. In the age of generative AI, a person's intellectual identity is not a free resource to be harvested. It is property. And the owners are coming to collect.