Sources & methodology

This is our own claim ledger. Before we put a number on our homepage, we checked it the way we’d check an AI answer: back to the original source, exact wording, sample and limits.

The figures we use

Menlo Ventures estimates that 2 billion people use AI worldwide. (2026)

Source
2026: The State of Consumer AI. Menlo Ventures. September 16, 2026. menlovc.com: 2026: The State of Consumer AI →
Exact original wording
“Consumer AI reached two billion users in 2026, roughly 24% of the global population according to Menlo Ventures’ estimates.”
“The number of AI users globally grew just 11% this year, from 1.8 billion to 2.0 billion people, while consumer spend more than tripled.”
Sample and population
The underlying survey covered 5,067 U.S. adults, run with Morning Consult in July 2026, nationally representative and weighted. No one outside the US was surveyed for this figure.
Definitions
The report doesn’t define a global “user” by how often they use AI. The survey asked about use “over the past six months.” Its market-sizing note says: “While the survey’s usage and spending figures are reflective of U.S. adults, the global estimate is adjusted downwards to account for differences in internet access, age distribution, and regional adoption and spend rates.”
Limitations
  • It’s an estimate extrapolated from a US survey, not a global count.
  • There’s no daily, weekly or monthly definition, so “use AI” can include occasional use.
  • The publisher is a venture investor whose portfolio includes AI companies.

In a 2026 survey of US and UK consumers, 7 in 10 AI chatbot users who felt responsible for checking said they check answers only sometimes, rarely or never. (F-Secure)

Source
AI chatbots: navigating distrust and usefulness in everyday life. F-Secure. Fielded April 2026, published September 2026. f-secure.com: Consumer AI survey, April 2026 →
Exact original wording
“Almost everyone says checking is their responsibility. Yet 70.1% check only sometimes, rarely, or never.”
“70.1% report checking only sometimes, rarely, or never, despite 89.5% assigning themselves full or partial responsibility for verification.”
“Regularly check 29.9%. Only sometimes 34.6%. Rarely or never 35.5%.”
Sample and population
1,500 consumers across the United States and United Kingdom (750 each), surveyed April 2026. AI users answering: n=1,045 to 1,165 depending on the question. The page’s chart gives the base for this figure as “n=933 who claim responsibility,” so it describes users who said checking is fully or mostly their responsibility, not every user.
Definitions
“AI chatbot” is defined on the page as “any software program that simulates human conversation.” “Regularly check” and “only sometimes” aren’t defined any further. The page calls respondents consumers and doesn’t state an age floor.
Limitations
  • It’s self-reported, and self-reports likely overstate checking.
  • It’s vendor-run: F-Secure sells consumer security products.
  • It covers the US and UK only, combined.
  • The figure includes “sometimes,” so it isn’t the share who don’t check at all.
  • The page’s summary describes the rarely-or-never group as checking “responses that sound confident,” while the chart says “answers.”

What “Most take its word for it” means

“Most take its word for it” summarizes survey findings that most users don’t regularly check AI answers:

  • F-Secure, US/UK, 2026: 70.1% of users who felt responsible for checking report checking only sometimes, rarely or never. 29.9% regularly check.
  • University of Melbourne & KPMG, 47 countries, 2025: 66% of employees using AI at work have relied on its output without evaluating it at least once. The report’s Figure 44 splits that 66% into “Rarely” 24% and “Sometimes to very often” 42%; 34% said “Never.”

A survey that points the other way: Clear Spark Digital, US, April 2026 (2,076 AI users on an opt-in Prolific panel): 17.4% always verify, but 54% verify “always” or “often,” 33.8% “sometimes,” and about 12% rarely or never. On this survey’s wording, most users say they check often, so we don’t count it as support for the headline.

Why people don’t check

From the same F-Secure survey: what stops users of AI chatbots from checking the answers. n=1,045, and people could choose more than one.

No reliable source to check against
37.5%
Answer already good enough
34.4%
Checking takes too long
19.2%
Don’t know how to check
15.4%

The chart has one more row, which we left out because it isn’t a barrier: “Can check easily,” 25.6%.

Figures we don’t use, and why

“92% don’t check their AI answers”

The survey measured clicking source links in Google AI Overviews.

Exploding Topics’ survey of 1,115 web users found “Just 7.71% report always following the links provided by AI Overviews.” Not clicking one link isn’t the same as not checking an answer.

explodingtopics.com: The AI Trust Gap →

“91% of users don’t fact-check”

Anthropic counted conversations without visible in-chat checking, and noted checking may happen elsewhere.

The AI Fluency Index found “checks facts and claims that matter” in 8.7% of 9,830 Claude.ai conversations. It counts conversations, not users, and it says “Users might demonstrate fluency behaviors mentally (such as fact-checking Claude’s claims against their own knowledge) without expressing these behaviors in conversation.”

Anthropic: The AI Fluency Index →

What we’ve tested

Outside studies describe the problem. They don’t measure Xyntherium. These are our own results, each with its limits.

Help Center Check held-out run

One run on November 2, 2026, on a frozen set: 3 of 89 labeled breaks missed (3.4%, 95% interval 1.2–9.4%). That’s under our 5% bar on the point estimate, not on the top of the interval. The set is spent.

About Help Center Check →