Google’s AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy
Citation. Iscenko, Z., Strand, S., Chen, Y., Aimard, G., Codreanu, M., Sampathkumar, V., Imas, A., Jacobs, J., Muiruri, E., Mateos-Garcia, J., Ng, J. J., Javed, S., Martin, J., Ajmeri, O., Calin, D., Kim, A., Curto Millet, F., & Manyika, J. (2026). Google’s AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy. SSRN. https://doi.org/10.2139/ssrn.7169260
Summary. This Google/Google DeepMind working paper introduces ATLAS, a descriptive study of 14.65 million de-identified Gemini App, AI Mode, and unpaid Gemini API interactions sampled April 6–19, 2026 and mapped to occupational, task, household-activity, geographic, and language taxonomies. It finds broad but shallow observed workplace use: Gemini appears across occupations covering 88% of US employment, while the median occupation with observed use has use on 21% of its tasks. Most observed non-routine cognitive use is classified as collaborative drafting, review, ideation, or information retrieval/learning rather than end-to-end automation.
Key claims
- Observed workplace Gemini use spans 68% of occupations globally, covering just above 88% of US employment, but only 3% of occupations show use on more than 75% of their constituent tasks.
- In the authors’ intent classifier, end-to-end automation accounts for less than 10% of observed non-routine cognitive-work conversations; partial drafting/generation, review/refinement, ideation/strategy, and information retrieval/learning dominate.
- A 1% increase in an occupation’s median earnings is associated with more than a 2.5% increase in AI-use intensity in the US data, even after the authors control for educational attainment; this descriptive gradient raises an adoption-inequality concern rather than establishing a causal effect.
- ATLAS’s automated task and activity labels received high human-expert reasonableness ratings at broad taxonomy levels, but granular task assignments are less certain and the report advises against over-relying on hyper-specific task classifications.
Evidence & limitations
- The study is a descriptive, observational analysis of 14,653,926 de-identified interactions from April 6–19, 2026, processed through an automated privacy-preserving summarization, clustering, and taxonomy-assignment pipeline and evaluated with synthetic-data, rater-agreement, and human-approval exercises.
- The sample excludes paid Gemini API content and several Google AI surfaces, cannot observe non-users or activity outside Gemini, and measures interactions rather than realized task completion, productivity, wages, employment, or causal effects.
- The paper is a July 2026 SSRN working paper from Google and Google DeepMind; its proprietary behavioral data offer unusually broad descriptive coverage but require independent replication and should not be read as an estimate of total AI use or workforce impact.