In July 2026, Google published one of the largest looks yet at how people actually use conversational AI: AI & the Economy: ATLAS v1.0, built on 14.65 million de-identified interactions sampled from the Gemini app, AI Mode, and the Gemini API between 6 and 19 April 2026, spanning more than 800 occupations, 150 countries, and 140 languages.
The report is full of findings about work. The most revealing number has nothing to do with work at all.
What does ATLAS say about how people use AI?
Over 86% of AI conversations happen outside of a job. And when you break those conversations down by activity, one category lands far above where anyone plotting "AI at work" would have put it.
| Activity | Share of all conversations |
|---|---|
| Socializing, relaxing & leisure | 26.4% |
| Education | 20.7% |
| Work & work-related | 13.5% |
| Household activities | 11.0% |
| Personal care | 9.0% |
| Other | 6.6% |
| Consumer purchases | 5.2% |
| Professional & personal care services | 4.8% |
| Traveling | 2.7% |
Education is 20.7% of everything people bring to these tools — roughly 24% of all non-work conversation, second only to socializing and leisure, and more than the entire work category. People are spending more of their AI on learning than on their jobs.
Why the education number is bigger than it looks
Share alone understates it. Google also compared how much AI assistance each activity attracts against how much time people actually spend on that activity in daily life. Most categories track: the more hours something takes, the more people ask about it. Eating and drinking, for instance, consumes about 18× more time than it does AI conversation.
Education breaks the pattern in the other direction. People turn to AI for learning 5.8× more than their time budgets would predict.
And it isn't alone at the top. The activities where AI is most over-represented relative to time spent are the ones that are hardest to face alone:
- Navigating government services and civic obligations — almost 20×
- Professional and personal care services (legal, medical, financial) — more than 7×
- Education — 5.8×
- Consumer purchases — nearly 3×
About half of those medical, legal, financial, and government conversations arrive outside working hours — at night, early in the morning, on weekends.
Read together, they tell one story. People don't only reach for AI to save time on what they already do a lot of. They reach for it to close a gap in understanding — to make sense of something confusing, intimidating, or hard to access on their own. Learning isn't a side effect of AI use. For a very large share of people, it's the point.
So why is software engineering the exception?
Hold that finding next to another one from the same report. Financial and investment analysts, market research analysts, software developers, and network and computer systems administrators are among the occupations most over-represented in AI use relative to their share of employment. And in the small 3% of occupations that use AI for more than three-quarters of their tasks, the first one listed is software quality assurance analysts and testers.
Coding is where AI has penetrated furthest into the actual work — not just drafting and summarizing, but generating, modifying, and inspecting code faster than a person can read it. Is engineering intelligence growing from this new scale of output, or just lines of code? That is a core question we ask at Cartara. Our goal is to lead the way in AI assisted development and help users understand AI generated code.
Everywhere else, people use AI to understand more. In software engineering, AI increasingly does the thing itself — and understanding is exactly what gets skipped.
ATLAS's own headline is that AI assists far more than it automates: fewer than 10% of interactions fully automate a task, and in a typical job AI is used for only ~21% of tasks. That is a real check on any "AI does everything now" claim, and we take it seriously.
Two things to hold alongside it. First, ATLAS samples the Gemini app, AI Mode, and the Gemini API — it barely observes agentic coding tools, which is precisely where end-to-end generation is concentrated. Second, assistance is not the same as comprehension. Code that arrives through a collaborative back-and-forth still arrives faster than a mental model forms.
What is knowledge debt?
When an agent writes a function end to end, you get the output. What you don't get is the thing that used to come bundled with it: the hour of struggling through the implementation, the dead ends, the moment the design finally clicked. That struggle was never the inefficiency. It was where the learning lived. Strip it out and you keep the velocity but lose the comprehension.
That's knowledge debt, and it compounds quietly:
- Changes get accepted without being understood.
- Junior engineers miss the reps that used to turn them into senior ones.
- Teams ship systems that only the agent — and perhaps no human — can fully explain.
Your repository remembers every line the agent wrote. The question is whether your team does. You can't debug, extend, review, or defend code you don't understand, no matter how fast it shipped.
Cartara: the learning layer for AI-native engineering
Cartara exists to resolve that tension. It's the learning layer that sits alongside your AI coding and turns each session into two outputs instead of one: the working change, and a durable increase in what you actually understand.
It works from evidence, not claims. Cartara analyzes what genuinely changed in your repository — your real diffs — never what an agent says it did. Because it reads the repository rather than the assistant, it doesn't much matter which tool wrote the code. Our goal is to build a foundational level of knowledge so engineering intelligence is an outcome from working with AI agents, not just having AI work for you.
From those diffs it surfaces the three to seven engineering concepts that actually mattered in the session, explains each one in the context of your own code, and checks your understanding with a short, three-minute knowledge check tied to what you just shipped. Results feed a deliberately conservative mastery model — exposure to a concept is not the same as mastering it, and Cartara refuses to pretend otherwise. Over weeks and months, a personal knowledge map shows where your understanding is deepening and where the gaps are.
The payoff scales past the individual. For engineering leaders, Cartara turns comprehension into something visible: aggregate, privacy-respecting insight into where knowledge is concentrated and where a single departure would leave a system unexplained. Not surveillance, and not vanity metrics — a real read on team resilience.
For classrooms, where the report's education signal is loudest, Cartara for Education lets students build real software with AI without offloading the learning that is the entire reason they're there, and gives teachers a per-student view of who actually understands what they turned in.
Turn your last coding session into concepts you can prove you understand
Cartara reads your real diffs, surfaces what mattered, and checks whether it landed.
What the data proves — and what it doesn't
Be precise about the limits, because this is where commentary on reports like ATLAS usually overreaches.
What ATLAS establishes: people use AI to learn at a rate that outstrips almost everything else they use it for, and that appetite is strongest exactly where a task is confusing, high-stakes, or hard to access alone. That is no longer speculative.
What it doesn't establish: that the same appetite survives the trip to work. ATLAS measures learning in personal contexts; nothing in it says a senior engineer merging an agent's pull request at 4pm on a Thursday wants a concept explained. The inference that they do — that the instinct is human rather than contextual, and that it's currently unserved because coding tools are built to produce code rather than comprehension — is the bet Cartara makes. We think the evidence points that way. We're not going to pretend the report says it outright.
Build with AI, and actually learn what you built — so that when the agent ships your software, you're the one who comes away smarter.
Frequently asked questions
Data source: Google, "AI & the Economy: ATLAS v1.0" (July 2026). Figures cited are Google's; the chart above is our own rendering of the report's Figure 10.