Google Says AI Isn’t Taking Your Job. Its Own Data Quietly Says More.
In a nutshell
There is a war of narratives underway about whether artificial intelligence will destroy jobs or merely assist with them, and it is being fought with data by the same companies that profit from the answer. This week Google fired its biggest salvo yet. gafam.ai reads what the study says, what it conveniently does not, and why the whole contest should worry Europe more than either narrative admits.
What Google Published
On July 23, Google released the first edition of its AI & Economy ATLAS report. It is built from roughly 15 million aggregated and de-identified human-AI interactions across the Gemini App, AI Mode and the Gemini API, which together are used by more than a billion people every month. The dataset spans more than 150 countries, 140 languages, 800 occupations and 4,000 tasks, making it one of the largest empirical looks at real AI usage assembled to date.
The headline is deliberately undramatic. AI adoption is broad but shallow: AI use appears in around 68% of occupations, together representing roughly 90% of US employment, but within a typical job it is used for only about 21% of tasks. And crucially, fewer than 10% of workplace interactions attempt to fully automate a task; most are collaboration — ideation, information retrieval, troubleshooting, strategy and learning. One of the report's own economists, Scott Strand, summed up the intended takeaway bluntly: just because you are using AI does not mean it is going to automate your job.
That is a reassuring message. It is also, not coincidentally, a very convenient one for Google.
The Number Google Left Out
Here is where gafam.ai's discipline earns its place, because the most revealing finding is one Google did not put in its announcement. According to an analysis of the report's own tables, automation intent rose above 25% — more than a quarter — for routine cognitive tasks, a figure the company's announcement omitted while highlighting the sub-10% figure for non-routine cognitive work.
That is a significant omission. The comforting less-than-10% headline applies to complex, non-routine cognitive work — the harder-to-automate category. For the routine cognitive tasks that make up a large share of many administrative, clerical and support jobs, Google's own data shows automation intent more than twice as high, and the company chose not to feature it. The reassurance in the press release is real for some workers and considerably thinner for others, and the difference falls precisely along the line of who is most exposed.
Three Reasons to Read the Whole Study Sceptically
Beyond the omitted figure, three methodological limits matter, and to Google's partial credit the report acknowledges some of them.
First, ATLAS measures intent, not outcome. It classifies what users asked the AI to do, not whether the AI succeeded or how effective the result was. The report itself states its findings should not be considered definitive. A user asking Gemini to automate a task is not evidence the task was automated.
Second, it is Google's own neighbourhood. The data comes entirely from Google's own products and usage logs — an unmatched view of behaviour, but not a census of AI use. It is a map of one company's territory, drawn by that company.
Third, and most consequentially, the study excludes Google's enterprise tools — Gemini Enterprise and Workspace — because, as Axios reported, Google does not maintain comparable logs for them. That omission cuts against the reassuring headline: enterprise deployments are exactly where automation of structured, repetitive, administrative workflows is most likely to be concentrated. Measuring the consumer Gemini app and leaving out the enterprise suite is measuring AI's labour impact with the most automating part of the picture removed.
The Other Camp: Anthropic and OpenAI
The reason Google's framing is not neutral becomes clear when you place it against its rivals. Google's conclusions are directionally similar to earlier work from Anthropic and OpenAI, but with a telling divergence: Anthropic reported a materially higher level of automation. Anthropic's own economic research has found automation accounting for something in the range of 43 to 45% of usage, against a slim majority for augmentation — and that automation share has been rising over time.
More broadly, Anthropic and OpenAI have been the leading voices of the mass-displacement narrative — the warning that AI could eliminate large swathes of white-collar and entry-level work within a few years. Google's ATLAS is, in effect, the empirical counter-argument to that warning.
And here is the point gafam.ai insists on: neither camp is a disinterested referee, and their narratives track their commercial interests with uncomfortable precision. Anthropic and OpenAI, selling frontier capability and raising capital at extraordinary valuations, benefit from a story in which their models are powerful enough to replace human labour — power justifies price, urgency justifies investment, and even the safety framing depends on the models being formidable. Google, defending a consumer product used by a billion people and facing regulatory and political heat on multiple fronts, benefits from a story in which its AI is a helpful assistant that empowers workers rather than a threat that displaces them. The augmentation narrative soothes the public whose data trains the model and the regulators deciding its fate. Each company is measuring reality, and each is measuring it in the way that flatters its business.
The European Perspective
For Europe, the narrative war over AI and jobs is not an abstract academic dispute — it is the single most politically consequential question artificial intelligence poses, and Europe has almost no independent capacity to answer it.
The augmentation-versus-automation question determines the future of the European social model: labour protections, works councils, retraining policy, education reform, the entire settlement between workers and technology that Europe has spent a century building. And yet the evidence base for how AI actually affects work is being produced almost entirely by American companies with direct commercial stakes in the conclusion.
Google's ATLAS spans 150 countries, but its employment framing is American — 90% of US employment, mapped to US Bureau of Labor Statistics categories. Anthropic's and OpenAI's economic research is American. There is no comparable large-scale European study of how AI is reshaping European work, because there is no European frontier AI product with a billion users generating the usage data such a study requires. This is a form of dependency gafam.ai has not previously named: epistemic dependency.
Europe does not merely rely on American models, chips and platforms; it relies on American companies to tell it what those models are doing to European livelihoods. When the European Commission, national governments and unions debate how to regulate workplace AI — and recall that the EU AI Act classifies employment-related AI as high-risk, though it deferred those obligations to December 2027 — they will do so using data curated and framed by the very firms whose products are under scrutiny, and whose commercial interest in the answer is direct. The mature European response is not to pick Google's reassuring story or Anthropic's alarming one, but to build the independent European capacity to measure AI's labour impact for itself: publicly funded, methodologically transparent, drawing on European labour data and European deployment, answerable to European citizens rather than to American shareholders.
A continent that lets others measure what is happening to its own workers has surrendered not just the tools of the AI economy but the ability to see clearly what that economy is doing to it. Europe regulates AI's effect on work as high-risk. It cannot yet independently observe it. gafam.ai will be watching.
We are not first. We are right.
SOURCES
— Google (The Keyword / blog): Understanding the AI economy: the first ATLAS report
— Axios: Google study finds broad AI use, but little evidence of job automation
— Implicator.ai: Google Finds Automation Intent Above 25% in Routine Work — a figure the announcement left out
— Silicon Snark: Google Mapped 15 Million AI Conversations — a map of Google's own neighbourhood, enterprise excluded
— Google ATLAS v1.0 (research PDF): Mapping Gemini Usage in the Economy — augmentation vs automation, Anthropic comparison
🔒 This analysis is for GAFAM Intelligence members only.
Already a member? Log in here