Google's own data on 15M Gemini chats shows AI is barely automating real jobs

Turns out the robot apocalypse is running late. A new study from Google Research combed through 15 million anonymized interactions with Gemini and found that, despite the breathless hype, AI is not about to gut the white-collar workforce. The paper, introducing their ‘AI & Economy ATLAS,’ is a rare moment of sobriety from a company pouring billions into the technology. It found that use ‘remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope.’

This isn’t a survey asking people how they feel. Researchers used an automated classifier to map real Gemini prompts to specific occupations in the Bureau of Labor Statistics’ database. Human reviewers then verified the classifications. It’s about as close to a candid X-ray of actual workplace usage as you can get, and the picture isn’t one of autonomous agents replacing workers. It’s a story of people using AI as a brainstorming partner or a search engine, not a replacement for their own judgment.

Unsurprisingly, the data showed who’s actually kicking the tires on this tech. White-collar roles in tech, finance, and creative fields were overrepresented. Software developers, financial analysts, and sysadmins are the heaviest users. Meanwhile, you could hear crickets from the prompts of salespeople, transportation workers, and food service staff, whose professions were dramatically underrepresented in the data relative to their share of the US economy.

The study’s blunt conclusion is bracing for anyone who predicted imminent mass displacement. ‘We [did] not find evidence… to support the claims that AI is about to cause massive automation and displacement of white-collar work,’ the authors write. While that will be cold comfort to anyone who loses their job to a specific AI tool, the aggregate data suggests we’re in for a period of messy, uneven augmentation—not a clean break with human labor. The real risk might not be that AI outsmarts us, but that we get bored of it before we figure out what it’s actually good for.