Google’s 15 Million AI Interactions Suggest Workers Are Not Automating Themselves Away

Lead

The public debate around generative AI often jumps straight to replacement: if models can write, code and reason, are office workers teaching machines to take their jobs? A new study from Google Research offers a more grounded view. By examining 15 million anonymized interactions across the Gemini App, Google’s AI Mode and the Gemini API, the team behind the AI & Economy ATLAS looked at how people are actually using AI at work.

Their conclusion is more measured than the hype cycle suggests: AI is being used across many occupations, but its role is still mostly shallow, partial and collaborative.

Key points

• Adoption does not equal automation. The researchers mapped work-related AI interactions to US Bureau of Labor Statistics occupational categories and O*NET task descriptions. They found substantial usage, but little evidence that entire jobs are being handed over to AI from end to end.
• Usage is concentrated in white-collar work. Financial and market analysts, software developers and systems administrators were overrepresented in Gemini work usage. Sales, transportation and food service roles were underrepresented, reflecting the advantage AI tools have in text-heavy and information-heavy workflows.
• Most jobs are only lightly touched. Across the O*NET task database, only 21 percent of work-related tasks met the study’s threshold for being considered “Gemini tasks.” For 29 percent of occupations, no relevant task reached that non-negligible usage threshold. Another 30 percent of occupations had meaningful Gemini use in less than a quarter of tracked tasks. Only 3 percent of occupations saw AI consulted for at least three-quarters of relevant tasks.
• Common uses are drafting, generation, retrieval and learning. In cognitive work, Gemini was most often used to create drafts, generate ideas, look up information or support learning. Interactions categorized as direct automation were a minority, even within tasks considered routine.
• Workers tend to delegate lower-expertise tasks. The study found that AI use clustered around lower-complexity cognitive work such as rewriting, translation, product specification drafting and review. The more expert and context-dependent parts of jobs were less likely to be fully offloaded to the model.

Why it matters

The study does not argue that AI has no labor-market impact. Instead, it draws an important distinction between what AI might theoretically do and what workers currently ask it to do. Today’s usage patterns suggest that generative AI functions more like a productivity layer inside existing workflows than a wholesale replacement for most occupations.

For companies, that implies AI strategy should focus less on abstract headcount reduction and more on redesigning workflows around specific bottlenecks: document drafting, information retrieval, summarization, translation and preliminary analysis. For workers, the signal is similarly nuanced. The risk is not simply that an occupation disappears overnight, but that its task mix changes. Routine cognitive fragments may shrink, while judgment, domain expertise, interpersonal coordination and non-routine problem solving become more valuable.

The picture could change. Future models may become more capable at high-expertise work, and AI-powered robots could eventually affect manual occupations more directly. But based on current Gemini interaction data, the dominant pattern is not workers automating themselves out of existence. It is workers using AI as a collaborator for selected parts of their jobs.

Source: Ars Technica AI