Competing Empirical Studies on AI's Actual Impact on Work

Two large-scale empirical studies — Google's ATLAS (15 million Gemini interactions) [1] and OpenAI's 'Work at the Frontier' report (800,000+ ChatGPT messages) [2] — published within days of each other in late July 2026, offer the most data-rich portraits yet of how AI is actually used at work. Both studies find no evidence of mass job displacement. Where they differ is in framing: Google emphasizes that AI use remains shallow and collaborative rather than replacing end-to-end task execution [1], while OpenAI documents a different kind of change — workers routinely using AI to perform tasks associated with occupations other than their own, with 43.5% of occupation-specific messages crossing job-category lines [2].

Workforce policy, hiring decisions, and retraining investment are all being shaped by predictions about AI-driven displacement. Studies of this scale — from the companies whose tools are being studied — are among the few empirical inputs available to policymakers. Their convergence on the 'no mass displacement yet' finding may reduce near-term alarm, but the task-crossover pattern OpenAI identifies suggests occupational boundaries are already shifting in ways that standard labor statistics won't capture for some time.

In the span of roughly a week in late July 2026, Google and OpenAI each published major empirical studies of how workers actually use AI tools. Both drew on tens of millions or hundreds of thousands of real interactions, mapped them against Bureau of Labor Statistics occupational classifications and O*NET work-activity data, and arrived at findings that complicate simple narratives about AI and employment — in different ways.

Google's ATLAS (Activity, Task, Landscape, and Adoption Study) analyzed 15 million anonymized interactions across Gemini App, Google's AI Mode, and the Gemini API, covering more than 800 occupations [1]. Its headline finding, as reported by Ars Technica's Kyle Orland, is that AI use 'remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope' — and that the data provides no support for claims that AI is about to cause massive displacement of white-collar work [1]. The study found broad adoption across occupations, but depth of automation, not breadth of adoption, is what drives displacement risk, and on that dimension ATLAS found little.

OpenAI's Economic Research team reached a complementary but differently-framed conclusion in its 'Work at the Frontier' report. Analyzing over 800,000 work-related ChatGPT messages, the team found that 43.5% of occupation-specific messages involved tasks normally associated with a different occupation — a pattern they call 'task crossover' [2]. Customer experience workers (77%), designers (75%), and HR workers (69%) showed the highest rates of using AI for tasks outside their own job domain. Marketing tasks spread most broadly across other fields, appearing in 8.9% of messages from non-marketing workers. Workers in small organizations (2–5 seats) showed higher crossover rates than those at large firms (18.9% vs. 16.3%), which the researchers attribute to AI functioning as a generalist substitute where specialist colleagues are absent [2].

Together, the two studies describe a picture of AI adoption that is wide but (so far) structurally shallow on the displacement dimension, while producing meaningful change in what individual workers actually do day to day. The ATLAS study's 'no displacement evidence' finding and OpenAI's task-crossover finding are not contradictory — a worker using AI to draft a contract or debug code may not be replacing a lawyer or engineer if the task is incidental and the output is lightly used. But OpenAI's researchers argue that usage patterns of this kind may be leading indicators of occupational change that conventional labor statistics will detect only later [2], which is precisely what makes the current empirical record difficult to interpret definitively.