Grok is on the rocks

This is a column about AI. My boyfriend works at Anthropic. See my full ethics disclosure here.

Today, let's talk about a tumultuous week in the life of xAI's middling chatbot, Grok, and what it reveals about the state of AI competition at this moment in 2025.

On July 9, xAI released Grok 4, calling it "the most intelligent model in the world." The model beat rivals on a handful of leading benchmarks, including one known as Humanity's Last Exam, where its score of 25.4 percent edges out Gemini 2.5 Pro by 3.8 points. But it lagged rivals at practical tasks, including coding and creative writing, leading to credible charges that xAI had gamed the benchmarks.

No one remembers any of that, of course, because immediately upon release, Grok 4 began generating antisemitic comments and referring to itself as MechaHitler. It also seemed to search Elon Musk's posts before responding to shape its views. A week later, it added a not-safe-for-work anime companion — in an app that to this day is rated for children 12 years and older in the App Store.

Grok has become a popular fixture on X, where users routinely invoke it to dunk on people making false claims. But despite some legitimately impressive aspects of xAI as an operation — most notably the speed with which it built models that are nearly state of the art — Musk's constant tampering and obsession with anti-wokeness have resulted in a series of comical, unforced errors that have blunted the app's momentum.

In an effort to kick-start Grok's growth, last week the company made its new image-to-video feature, Imagine, free for all users. Musk said that, as a result, Imagine was "growing like wildfire."

Over the past few days, though, something appears to have doused the flame. The top free app in the App Store is OpenAI's ChatGPT — while Grok is a still respectable No. 5.

But where most would see healthy competition between two well-funded AI labs, Musk — as he so often does — saw conspiracy.