Language Log — Super Fakehuman Grammar Everything Advice
Summary: Mark Liberman documents Grammarly/Superhuman’s “expert review” feature, which impersonated living and dead writers without consent, and the legal and ethical backlash that followed.
Sources: Raw/Language Log » Super Fakehuman grammar everything advice.md
Last updated: 2026-05-07
What happened
After Grammarly rebranded as part of Superhuman, it launched an “expert review” feature offering users writing feedback attributed to named real individuals — including Stephen King, the late bell hooks, former Microsoft CPO Julie Brill, data science professor Mar Hicks, journalist Kara Swisher, and NYT journalist Julia Angwin. None of these people consented, were contacted, or had any involvement. A disclaimer buried in the product stated the references were “for informational purposes only and do not indicate any affiliation.”
Angwin described receiving suggestions supposedly from “Julia Angwin” that were “clunky” and “weirdly unspecific” — a deepfake of her professional judgment packaged as a paid service feature.
The legal response
Angwin became lead plaintiff in a class-action suit in the US District Court for the Southern District of New York, alleging violation of New York and California right of publicity statutes — century-old laws prohibiting use of a person’s name or image for commercial purposes without consent. At least 25 states have similar statutes.
Superhuman CEO Shishir Mehrotra disabled the feature and announced plans to give “experts real control over how they want to be represented — or not represented at all,” while stating the company “believes the legal claims are without merit.”
Why it matters beyond the lawsuit
Liberman frames the feature as Grammarly having burst the “plastic fetters of grammar” to consume the entire curriculum (Trivium + Quadrivium + philosophy and theology) — meaning an AI writing assistant originally scoped to correctness has expanded to impersonate human judgment across every domain. The irony: the company positions this as democratizing expertise while eliminating the actual experts.
This is a concrete instance of what magfrump-sculpted-interaction identifies as the interface-layer alignment problem: the chatbot format substitutes a plausible simulation of expertise for the real thing, and the user cannot tell the difference.
Connections
The identity appropriation connects to identity — Angwin’s phrasing is precise: “an A.I. company was selling a deepfake of my mind.” Writing voice is an identity asset, not merely a stylistic property.
ai-alignment-failure-modes includes interface-layer undermining of human judgment as a failure mode. Fake expert review is a commercial instance: users receive guidance calibrated to feel authoritative rather than to be accurate, and the attribution to a named human makes the simulation harder to challenge.
language-log-gigo is structurally parallel: in both cases, a plausible-looking output (a paper, a writing suggestion) is attributed to a credible source that had nothing to do with producing it.
taylor-ai-honesty-policy is the normative flip side: what would it look like for AI communication to be genuinely credible? The Grammarly feature is a case study in what happens when commercial incentives push in the opposite direction.