AI in daily life is mainstream, mundane and measurably narrow

Half of US adults use an AI chatbot, and a quarter of the country reaches for one every day. Those two numbers close the adoption argument. They come from a Pew Research Center survey of 5,119 US adults fielded 17 to 23 February 2026, with a margin of sampling error of plus or minus 1.6 percentage points. But what they don’t settle is what actually changed, which is the harder question.

Here’s the claim I’d defend: AI in daily life is now mainstream and overwhelmingly mundane, and the places where it measurably changed an outcome rather than a habit are narrow enough to list on one hand. Dinner recipes at 6pm, tax questions on 14 April, a search page you no longer click through. One genuine exception, in road safety, where somebody publishes crash counts. Everything else is either a convenience, a projection, or a promise that hasn’t survived measurement yet, though plenty of people still quote it as though it had.

The rest of this piece is the evidence, with the sample sizes and the seams in the data left visible.

Half of US adults are in, and a quarter are in daily

Pew’s topline is blunt. Asked “Do you ever use an artificial intelligence (AI) chatbot like ChatGPT, Gemini or Copilot?”, 49% said yes and 51% said no. Frequency is where the habit shows, because ever-used and daily-used are different products. Among all adults, 4% use a chatbot almost constantly, 12% several times a day and 8% about once a day, which adds to roughly a quarter of the country on a daily cadence. That’s my arithmetic on Pew’s published shares.

How often US adults use an AI chatbot	Share of all adults
Almost constantly	4%
Several times a day	12%
About once a day	8%
Several times a week	10%
Less often	15%
Does not use AI chatbots	51%

The spread by age is wider than the headline. Two thirds of adults aged 18 to 29 use chatbots, against 23% of those 65 and over. ChatGPT alone reaches 44% of adults, up from 34% a year earlier, with Gemini second at about a quarter. And devices tell a quieter version of the same story, though nobody markets them as AI any more: 37% have a smartwatch and 35% a smart speaker, while roughly one in ten report a robot vacuum or an AI thermostat.

One seam matters before anyone quotes the growth rate. Pew’s own footnote records that before 2026 the question was worded differently and was asked only of people who had heard at least a little about chatbots. In the August 2024 wave, 28% had heard nothing at all. So the move from 33% to 49% spans a methodology change as well as two years of adoption, but Pew says so on the page rather than burying it, which is more than most adoption charts manage.

What people actually ask for is dinner, sleep and tax

The best public read on what gets typed comes from inside a lab, because that’s the only place the conversations sit. In How People Use ChatGPT, an NBER working paper issued in September 2025, seven authors classified representative samples of conversations from November 2022 to July 2025. Non-work messages grew from 53% to more than 70% of all usage. Three categories, Practical Guidance, Seeking Information and Writing, accounted for roughly 80% of conversations. And by July 2025 the product reached about 10% of the world’s adult population.

That matters because it inverts the pitch. The marketing was about work. But the measured majority of messages isn’t work at all, and the single largest work use is still writing, which is the task chatbots were already good at in 2023.

Anthropic’s Economic Index report on cadences, covering conversations sampled between 10 April and 10 June 2026, puts a clock on it. The share of conversations classified as personal use rises from around 35% on weekdays to just under 50% at weekends. Hour by hour, people ask for news at 7am local time, recipe requests run 2.3 times more frequent at 6pm than their daily average, and sleep advice peaks around 5am. Still, the sharpest spike in the whole sample was fiscal rather than domestic.

On April 14, tax-related clusters were eight times as common as on the average day in May and remained about as high on April 15. On April 16, they dropped sharply.

Anthropic Economic Index report: Cadences, June 2026

Read that as a measurement of what AI replaced, and it’s modest. It replaced a search, a forum thread, a phone call to someone who did their taxes last year. That’s useful, repeatedly, at the margin of an evening, but it isn’t a rearranged life.

The clearest measured change in daily life isn’t that AI does your job. It’s that it answers the question you used to type into a search box, and the search box stopped sending you anywhere.

Rundowns AI

Search changed under you without being announced

This is the change most people absorbed without choosing it. Pew ran a metered panel of 900 US adults and logged every visit to a search engine during March 2025, covering 68,879 unique Google searches. An AI summary appeared on 12,593 of them, around one in five. And the behavioural gap on either side of that summary is large.

Users who met an AI summary clicked a traditional result in 8% of visits, against 15% when no summary appeared, so roughly half as often. They clicked a link inside the summary on just 1% of visits. And they ended the browsing session outright on 26% of summary pages, against 16% of ordinary results pages. Median summary length was 67 words, so the answer that replaced the click was short.

By February 2026, 60% of US adults told Pew they read those summaries, 30% said they don’t and 10% weren’t sure. The catch is that reading an answer and reaching for it deliberately are different behaviours. The Reuters Institute’s Digital News Report 2026, built on 97,520 respondents across 48 markets with roughly 2,000 per market, found the deliberate version growing unevenly.

Weekly use of AI chatbots for news rose from 7% to 10% globally, and the report notes the youngest age group’s 17% is three times the oldest group’s 5%. But that growth came from parts of Asia, Africa, Latin America and southern and eastern Europe. The report states flatly that several markets including the USA, the UK, France and Germany “reported no increase in AI chatbot use for news over the year”. Pew’s separate US figure is consistent: 13% of adults say they use chatbots for getting news. So the passive change in search is large while the active change in news habits has stalled in the richest markets, which is roughly the gap we hit when we looked at how AI benchmarks get read, where a headline number moves far more money than the sample it rests on.

The one change with injury counts attached

If you want a part of daily life where AI changed an outcome and somebody published the denominator, it’s the road. Waymo’s safety hub reports 271.3M rider-only miles driven without a human driver through June 2026, split across Phoenix at 92.121M, the San Francisco Bay Area at 82.421M, Los Angeles at 67.098M, Austin at 21.064M and Atlanta at 8.624M. And its sustainability page puts current volume at “500K+ EV trips per week”, which is the throughput behind those miles.

Waymo Driver vs human benchmark, same distance, same cities	Reduction	Crashes avoided
Serious injury or worse crashes	95%	55 fewer
Pedestrian crashes with injuries	93%	87 fewer
Cyclist crashes with injuries	86%	56 fewer
Injury-causing crashes	82%	841 fewer
Airbag deployment in any vehicle crash	82%	357 fewer

Now the counter-reading, because it’s the strongest objection to using this table at all. The benchmark is constructed by Waymo, from Waymo’s own methodology, comparing Waymo’s miles to a modelled human rate for surface streets in five metro areas Waymo chose. The company says the data updates on the timelines of NHTSA’s Standing General Order reporting. So that’s voluntary transparency rather than independent audit, and the distinction matters because nobody outside the company can rebuild the denominator.

Even so, it’s the only consumer-facing AI product I found that publishes a rate, a denominator and a comparison group at all. The chatbots publish user counts. But a user count measures reach, not benefit, and the gap between those two words is where most AI claims live.

Where the promise still outruns the measurement

The productivity story is the one that keeps failing its own test, which is why it belongs here rather than in the section above. METR ran a randomised trial with 16 experienced open-source developers across 246 real issues in repositories they already maintained. With AI tools allowed, they took 19% longer to finish. Before starting, the same developers expected a 24% speedup. Afterwards, they still believed they’d been sped up by about 20%.

METR is careful about what that does and doesn’t show. The authors state plainly that their result isn’t evidence AI fails to speed up most developers, isn’t generalisable beyond their cohort and repositories, and is a snapshot of early-2025 tooling. Take all those caveats at full strength and one finding survives: self-reported time savings and measured time savings pointed in opposite directions in the same room. By my arithmetic that’s a 39 percentage point spread between belief and stopwatch. That’s the same gap we kept hitting when we checked what the hiring data shows about AI and jobs, where the aggregate employment series stays flat while one narrow cohort moves.

Healthcare runs the same pattern with better paperwork and worse evidence. The FDA maintains a list of AI-enabled medical devices it has authorised. I pulled the page on 12 October 2026 and counted 1,614 entries, each with its own submission number, the most recent decided on 29 June 2026. That count is mine, because the FDA publishes no total, and it warns in its own words that “the list is not a comprehensive resource of AI-enabled medical devices” since entries were identified largely by AI-related terms in summary descriptions.

And what’s in those summaries is thinner than clearance implies. A team led by Ahmed M Salih screened 1,105 FDA device summary reports published between 2021 and 2025 and analysed 519, scoring each against the six FUTURE-AI principles: fairness, universality, traceability, usability, robustness and explainability. Their paper, submitted on 4 July 2026, reports that 24.7% provided no evidence for any principle and none documented all six. Robustness appeared most often at 57.6%, traceability in 8.3% and explainability in 3.5%. Neither the year of clearance nor the clinical domain predicted better reporting.

So a device can be in your clinic, reading your scan, and the public record won’t tell an outside reader how it explains itself. That isn’t a scandal. It’s a measurement vacuum, which is the thing this whole piece keeps running into.

The bill arrives on the electricity meter

One measured change in daily life has nothing to do with whether you use a chatbot, so it lands on everyone. The IEA reported on 16 April 2026 that electricity demand from data centres “soared by 17% in 2025, and that of AI-focused data centres climbed even faster”, well outpacing growth in global electricity demand of 3%. That’s recorded consumption, so it belongs in this piece.

The forecast sits next to it, though it deserves a different label. The IEA’s Energy and AI base case puts data centre consumption at around 415 TWh in 2024, about 1.5% of world electricity consumption, and set to “more than double to around 945 TWh by 2030”. That’s just under 3% of total global electricity consumption in 2030, growing at roughly 15% a year. The same executive summary puts data centres at “around one-tenth of global electricity demand growth to 2030, less than the share from industrial motors, air conditioning in homes and offices, or electric vehicles”. Both halves of that are worth carrying, because the sector grows fast but stays small.

The honest summary, and what would change it

Put the measurements side by side and daily life in 2026 looks like this. Roughly half of US adults consult a chatbot, a quarter do it daily, and most of what they ask isn’t work. Search answers more and refers less.

Road injuries in five metros fell against a vendor-built benchmark. Developer productivity, tested properly once, went backwards while feeling better. Medical AI is authorised faster than it’s documented. And the electricity bill for all of it is real, up 17% in 2025, still a rounding error on the grid.

Nothing in that list is trivial. It’s also not the transformation the launch posts described, and the honest framing is that we’re nearly four years on from ChatGPT’s public debut with one well-measured outcome improvement to show for it.

Four things would change my mind, and I’d rather name them now than argue about them later. First, a Pew wave where daily use crosses 40% and the reported uses shift out of practical guidance and information lookup into something people would pay to keep. Second, a METR-style randomised replication on 2026 agents that finds a real speedup on real repositories. Third, a robotaxi crash comparison published by a regulator rather than by the operator, holding up at the same magnitude. Fourth, FDA device summaries where explainability runs well above single digits.

Until then, the measured story is adoption without much displacement, which is also roughly what we concluded about AI at work being a hiring story. If you want to know what AI changed in your own daily life, the test isn’t how often you open it. It’s whether you’d notice its absence by Thursday.