Google: Despite the AI Hype, Workers Aren’t Automating Themselves Away

Researchers at Google Research looked at how employees actually use the Gemini neural network, and found no support for the popular claim that artificial intelligence is about to trigger mass automation and push white-collar workers out of the labor market. The work, presented last week, draws on the AI & Economy ATLAS project — an analysis of 15 million anonymized interactions with the Gemini app, AI Mode, and the Gemini API.

To gauge the scale, the authors ran the work-related queries through an automated classifier built on the U.S. Bureau of Labor Statistics occupational classifications and the detailed O*NET database of work tasks. Manual checking confirmed the method’s reliability.

Usage exists, but it’s shallow

The heaviest users of Gemini are financial and market analysts, software developers, and systems administrators. The lightest are salespeople, transportation workers, and food-service staff. Yet even where AI is used actively, the integration remains, in the researchers’ words, “shallow and predominantly collaborative,” and end-to-end task automation is limited.

The numbers speak for themselves. Only 21% of all work tasks in the O*NET database drew a meaningful volume of Gemini queries. For 29% of occupations, not a single relevant task cleared this threshold, and for another 30% fewer than a quarter of tasks received significant AI help. And in only 3% of occupations — among them software testers, HR specialists, and records-management specialists — was Gemini regularly enlisted for three-quarters of tasks.

AI gets the routine, not the complex

The overwhelming majority of queries (86%) go to cognitive tasks — that is, “mental” work. But this mostly means drafts, brainstorming, and information lookup, rather than full automation.

Curiously, people entrust the model with precisely the tasks that require the least expertise: rewriting texts in other languages, and drafting and checking specifications. The hardest parts of their work, people keep for themselves.

Blue-collar workers weren’t overlooked either. Industrial-equipment mechanics use Gemini to analyze test results and error messages, while auto mechanics use it to inspect components and diagnose wiring; such users, more often than others, upload photos to the model instead of text.

The researchers’ overall conclusion: workers are not “automating themselves out of existence,” but complementing their labor — offloading routine cognitive operations onto AI and collaborating with it on non-standard tasks. The picture may change if future models learn to handle complex work, or if AI robots master manual labor. For now, the data indicate that the value of human skills in the non-standard tasks that dominate most professions is only rising.

Source: Ars Technica