A year ago, the New York Times Magazine ran an issue on artificial intelligence with the cover line “Everyone Is Using A.I. for Everything. Is That Bad?” It was a transcript from the Hard Fork podcast, and it assumed two things that have since turned out to be false. The first assumption: once you’ve tried AI, you use it “for everything.” The second: AI has gotten so good that despite any misgivings, “everyone is using” it. DuckDuckGo founder Gabriel Weinberg spent the year since then gathering the usage data the NYT piece did not, and published his analysis on Saturday. The numbers are blunt. About 70 percent of the US working-age population is not using AI at all. Gen Z AI adoption has stalled. Roughly a fifth of the population has never used AI. The “everyone is using AI for everything” framing is not the data. It is the marketing.

Weinberg’s analysis is the most useful piece of usage data published this year, because it is built from three independent sources — Gallup’s 2026 survey, Microsoft’s new US AI Diffusion report (based on anonymised telemetry from ChatGPT, Gemini, Claude, Copilot, and others), and a 2025 Datos study of desktop device behaviour — and all three sources say the same thing. AI is a technology with high awareness, low active use, and a long tail of resistance. That is not what “everyone is using it for everything” looks like.

🔍 THE BOTTOM LINE

Three independent sources now agree: the AI usage curve is steep at the top, flat through the middle, and resistant at the bottom. About 70 percent of US working-age adults are not using AI at all. Roughly 30 percent are. Of that 30 percent, a third use it only monthly or less. The “everyone is using AI for everything” narrative is not supported by Gallup, Microsoft, or Datos. The technology is real, the adoption is uneven, and the gap between the marketing and the data is the gap between a bubble narrative and a real curve. For New Zealand, this is a useful corrective: the assumption that the entire economy is being remade by AI in real time is not the data, it is the framing the loudest actors in the market want you to internalise.

The Gallup Data

Gallup has been running the same AI usage question for two years. The 2025 and 2026 numbers are almost identical, which is the most important finding: the adoption curve has not been steep, it has been flat.

The “at least rarely” number includes anyone who has used AI once in the last year. It is a very low bar. The number that matters is the intersection of “uses AI monthly or less” and “never uses AI” — together, about 50 percent of the US adult population is using AI rarely, monthly, or never. The “uses AI regularly” number — meaning at least weekly, in a meaningful way — is somewhere around 15–20 percent, depending on how you define “meaningful.” That is a smaller active base than the NYT framing implies.

The anger number is the most striking. “Angry about AI” rose from 22 percent to 31 percent year-over-year, the largest single jump in any of the categories. The people who are paying attention to AI are increasingly negative about it. The people who are not paying attention are not using it. The “enthusiastic majority” that the AI-marketing complex likes to invoke does not exist in this data.

The Microsoft Data

Microsoft’s US AI Diffusion report, published in May 2026, draws on anonymised, aggregated telemetry from Microsoft Copilot plus third-party data integrations. The number Microsoft cites: “more than 30 percent of the US working-age population is using AI.” The qualifier that gets less attention: that 30 percent is defined as having at least 90 minutes of usage time per month with major AI services. So the active US user base is around 30 percent by Microsoft’s own measurement. The other 70 percent is not using AI in a way that Microsoft can detect.

Microsoft’s data is the most direct measurement we have of actual AI usage, because it is based on telemetry, not self-report. Self-reporting overstates usage (people over-claim to pollsters). Telemetry understates it slightly (it misses people who use AI tools that are not connected to the major platforms). The 30 percent number is the closest thing to ground truth we have on US AI adoption. The NYT framing of “everyone” implies something like 80–90 percent adoption. The real number is closer to 30 percent.

Microsoft’s blog post on the report says the usage rate increased by 3 percentage points from the end of 2025 to mid-2026. That is not a hockey stick. It is a slow, even, gradual increase. The technology is being adopted. It is not being adopted at the rate the headlines imply.

The Datos Data

The Datos study, from 2025, used desktop device behaviour as its measurement. It found that as of last June, 21 percent of desktop devices visited “AI Tools” 10 or more times per month, 62 percent visited 0 times, and 17 percent were in between. The 62 percent “visited 0 times” is the most important number in the entire dataset. The vast majority of desktop users did not visit an AI tool at all in the measurement period.

There is a complication: the Datos study only measures desktop, and AI usage is increasingly mobile. A user who only accesses AI through their phone would show up as 0 visits in the desktop data. Even with that caveat, the 21 percent “10+ visits per month” number is consistent with Microsoft’s 30 percent “at least 90 minutes per month” number, and Gallup’s roughly 20 percent “regular use” estimate. Three independent measurements using three different methodologies are converging on the same answer: about a fifth to a third of the population uses AI regularly. The rest do not.

What Gen Z Tells Us

The Pew Research survey on Gen Z is the most interesting data point in Weinberg’s analysis. Gen Z has the highest AI awareness of any generation. In theory, they should be the heaviest users. The actual data says the opposite: Gen Z AI adoption has all but stalled in the last year, with a meaningful percentage of the population using AI rarely or not at all. Weinberg cites Gallup’s year-over-year numbers: 79 percent of Gen Z used AI at least rarely in 2025, 81 percent in 2026. The “regular use” number has been flat. The “never use” number has been flat. The “anxious about” number has been flat.

This is the generation that grew up with the technology, has the lowest friction to access, and has the most upside from the productivity gains. If Gen Z is not adopting AI at a steep rate, the assumption that the next generation will close the gap through natural adoption is wrong. The adoption curve is generational, but it is also demographically and economically constrained. People with discretionary time, discretionary income, and a job that benefits from AI use are the ones adopting. People without those resources are not. The technology is not democratising on its own. It is being adopted by the people who can afford to.

The “Jagged Frontier” Confirmation

The data is also a quiet confirmation of the “jagged frontier” thesis that Hassabis put forward on Friday with his Einstein Test proposal. The model is smarter than the median user at most tasks, and the median user is not using it at all. The frontier of model capability is advancing fast. The frontier of user adoption is advancing slowly. The gap between the two is the jagged frontier. It is the same gap that produced the Deloitte finding earlier this year that 40 percent of enterprise AI deployments were abandoned within six months. The technology is real. The use cases are real. The adoption is real but bounded.

This is also a check on the AGI-imminent narrative. If 70 percent of the US working-age population is not using AI, the question of whether AI has reached human-level general intelligence is, for most people, a question about a technology they have not engaged with. The DeepMind paper on AGI to ASI and the Anthropic Fable 5 export control story are both about frontier capabilities that 70 percent of the population has not touched. The gap between the frontier and the median is the gap the research papers do not measure and the NYT framing does not acknowledge.

What It Means for New Zealand

The most useful version of this data for a New Zealand reader is the one that says: do not assume the entire economy is being remade in real time. About 30 percent of the US workforce is using AI regularly. New Zealand is smaller and structurally more conservative. The number here is probably 20–25 percent, with the same demographic skew (younger, urban, professional services, discretionary time). For the other 75–80 percent of the New Zealand workforce, AI is not yet a working tool. It is a story.

That has policy implications. The sovereign-AI framework being pushed at the G7 assumes that AI infrastructure is a national security priority on the same level as electricity, water, and roads. The data says 70 percent of the population is not yet using it. The sovereign-AI investment case is forward-looking. The sovereign-AI risk case is the same. Neither is wrong. Both are real. The right response is to invest in the local applied layer (data, fine-tuning, deployment, local-AI hardware) at the same time as investing in adoption support for the 70 percent who are not yet using the technology. The two investments are not in conflict. They are the same investment at different horizons.

The other implication is the consumer story. The narrative that “everyone is using AI for everything” drives consumer behaviour in ways that are out of step with the data. People who are not actually using AI are told they should be. People who are using AI are told they are behind. Both groups internalise an anxiety that is not supported by the data. The honest version is: a fifth of the population is using AI regularly and finding it useful, a third is using it occasionally, and almost half is not using it at all. The technology is real. The adoption is uneven. The hype is not.

⚠️ THE OTHER SIDE

Three honest caveats. First, the data is US-centric. Microsoft, Gallup, and Datos are all US or US-heavy. New Zealand usage patterns are likely similar in shape but lower in magnitude. We do not have a local Microsoft AI Diffusion equivalent. Second, “using AI” is hard to define. The 30 percent Microsoft number requires 90 minutes per month. The Gallup “regular use” number is a self-report. A New Zealand tradesperson who uses ChatGPT to draft a quote for a client once a quarter is using AI in a way that matters to their business but does not show up in either dataset. Third, the adoption curve is not flat — it is up and to the right. The 30 percent number in mid-2026 was 27 percent at the end of 2025. The trend is real. The question is whether the trend is a slow climb to 50 percent over five years or a steep climb to 80 percent over two. The data is consistent with either. The “jagged frontier” thesis says slow. The “AGI to ASI cascade” thesis says steep. The data will tell us which in the next 18 months.

❓ FREQUENTLY ASKED QUESTIONS

What did Gabriel Weinberg actually say? He published a Substack post on Saturday titled “No, everyone is not using AI for everything.” His central claim: the NYT framing of “everyone is using AI for everything” is supported by neither Gallup survey data, nor Microsoft telemetry, nor Datos desktop behaviour. The actual adoption curve is steep at the top, flat through the middle, and resistant at the bottom. Roughly 30 percent of the US working-age population uses AI in a measurable way. Roughly 70 percent does not.

What is the Microsoft AI Diffusion report? A May 2026 report from Microsoft, based on anonymised, aggregated telemetry from Copilot and other major AI services. Microsoft defines “using AI” as at least 90 minutes of usage time per month with ChatGPT, Gemini, Claude, Copilot, or another major platform. The number: 30 percent of the US working-age population, up 3 percentage points from the end of 2025. The 3-point year-over-year increase is consistent with Gallup’s near-flat numbers, not with the “AI is taking over” narrative.

What does Gen Z’s data show? Gen Z has the highest AI awareness of any generation, but their adoption has stalled in the last year. Roughly 21 percent of Gen Z never use AI. Roughly 32 percent use AI only monthly or less. The “regular use” number has been flat year-over-year. The assumption that the youngest generation is the heaviest user is not supported by the data.

Is the data US-only? Yes, primarily. Microsoft, Gallup, and Datos are all US or US-heavy. New Zealand-specific data is limited. The shape of the curve is probably similar in New Zealand, with a smaller magnitude (probably 20–25 percent active use, 75–80 percent not). The Microsoft report does include some international comparison data; New Zealand is not broken out separately.

Does this mean the AI bubble is bursting? No. The data shows the technology is being adopted, slowly, by about a third of the population. That is a real curve. It is not a bubble curve. Bubble curves look like the 1999 internet adoption rate, which went from 20 percent to 60 percent in three years. AI is at 30 percent and adding 3 percentage points per year. That is the curve of a real technology being adopted unevenly, not the curve of a speculative frenzy about to collapse. The honest read: AI is a real technology with a real adoption curve. The hype is the gap between the curve and the marketing.

What is the “jagged frontier”? The term Hassabis used in his Einstein Test proposal on Friday. The frontier of AI capability is advancing fast, but the frontier of user adoption is advancing slowly. The gap between the two is the jagged frontier. It is the gap between what AI can do and what the population is actually using it for. The Weinberg data confirms the jagged frontier thesis with usage data: the capability is real, the adoption is real but bounded, and the gap between the two is the most important variable to track in 2026.

What should a New Zealand policy reader take from this? The data is a check on the assumption that the entire economy is being remade in real time. The sovereign-AI investment case is forward-looking. The risk case is the same. The right response is to invest in the local applied layer — data, fine-tuning, deployment, local-AI hardware — at the same time as investing in adoption support for the 70 percent who are not yet using the technology. The two investments are the same investment at different horizons.

Is the adoption rate accelerating? Slightly. Microsoft’s data shows 3 percentage points year-over-year. Gallup shows near-flat adoption. The Datos desktop data is from 2025 and does not show acceleration. The honest read: the adoption rate is slow and steady, not exponential. The “AI is taking over the world in 18 months” framing is not supported by the data. The “AI is being adopted unevenly and will continue to be” framing is.