"Anecdotes Everywhere, Evidence Almost Nowhere" is a July 20, 2026 essay by Steve Newman published on Second Thoughts (secondthoughts.ai), subtitled "The state of AI in mid 2026, part 3." It is the third installment of a series whose first two parts covered the pace of AI progress and the factors driving it. It is an opinion piece and is treated here as an argument.
Its thesis is stated in two lines: "AI has had, with a few exceptions, very little clear impact on the world at large," and "Strap in! That may be about to change."
The measurement problem
Newman opens with Anthropic's revenue as an illustration of both halves. The most recent figure he cites places Anthropic's revenue at 0.037% of the global economy — negligible — while the growth behind it runs from $1B at the start of the previous year to $47B ARR, with a footnote noting a newer report at $60B. He observes that continued growth at that rate would have Anthropic subsuming world GDP by the end of 2029, which will not happen, but which marks AI as "at the threshold of global importance."
He accepts the San Francisco comparison of AI in 2026 to Covid in early 2020 — not yet visible in most places, but visibly coming — while noting the analogy is imperfect, since "predicting the course of a profoundly novel technology is harder than predicting the course of a respiratory disease."
On why impact is hard to discern, his argument is that the available statistics mostly do not mean much. He gives Census Bureau business-survey data showing about 18% of firms had adopted AI as of year-end 2025, and argues the figure "tells us exactly nothing," since it records firms checking a box saying they use AI "in any of [their] business functions" — a category spanning an A/B testing tool and replacing an entire workforce. Beyond that, macroeconomic measures are buffeted by the Iran war, continuing Covid-lockdown ripple effects, and education-policy changes, making attribution to AI difficult; he cites the New York Times' July 2, 2026 piece "A.I. Is Reshaping the Economy. Good Luck Measuring How," and notes that a decline in Stack Overflow question volume — a widely used AI-impact proxy — predates the chatbot era. He adds that AI's impacts are highly uneven, so dramatic anecdotes are of unknown representativeness, while global AI capacity grows at roughly 10× per year. See AI and Productivity.
The one unambiguous effect
Newman identifies the capital expenditure itself, not any use of the technology, as the clearest economic impact: "the titanic quantities of cash spent building data centers, filling them with chips, and providing them with electricity," with Nvidia now the world's most valuable firm and, quoting Anton Leicht, "shaping up to be the largest infrastructure push in human history."
He cites a St. Louis Fed estimate that AI-related investment contributed 0.97 percentage points to real GDP growth in the first three quarters of 2025 — against a typical whole-economy growth rate of about 2.5% per year — and reads AI investment as a primary driver of US economic growth, possibly crowding out investment elsewhere. He attaches a substantial caveat of his own: the figure sums all spending on information-processing equipment, software, R&D, and data-center construction, not only AI-related spending, so it works as an AI estimate only if non-AI spending in those categories was flat year over year.
He notes that some economists speculate AI added one or two percent to labor productivity in 2025, calls this "an outlier view," and observes that AI investment outside the US is much more modest, with hot spots in Taiwan and South Korea. See AI Bubble vs. Buildout — Synthesis.
Labor market
US unemployment has drifted up modestly over several years, with the upward trend beginning around ChatGPT's launch — a coincidence Newman declines to read as causal, listing the Russo-Ukrainian war, Fed rate hikes, tariffs, and the post-Covid tech-hiring hangover as competing explanations.
The strongest specific finding he cites is the Stanford "Canaries in the Coal Mine" study, which found employment 16% lower for early-career workers in the most AI-exposed occupations. He immediately bounds it: the affected group is narrow (about 7% of workers studied); no broad employment effect is visible even for software engineers; the age 20–24 unemployment rate is roughly unchanged since the AI boom began; and the 16% figure is relative to similar-age workers in less-exposed jobs after controls, with absolute employment dropping less than 10% against rising employment in other groups.
His stated position is deliberately unresolved: he does not have a clear idea what AI's impact on jobs is today; he is certain of enormous change over time; he guesses overall employment is likely to decrease; and he thinks it possible this is already occurring and hitting entry-level positions first. But "there is no clear signal of impact on overall employment, which means the impact cannot be very large so far." Visible impact appears in specific occupations — translation, freelance writing, stock illustration. He also notes second-order effects: job hunting made more miserable by mass-produced cover letters meeting AI resume screening, and AI enabling worker "emotional surveillance." See AI Labor Disruption.
Education
Newman opens the section with Zvi Mowshowitz's formulation: "If you want to use AI to learn, it is the best tool ever invented for learning. If you want to use AI to not learn, it is the best tool ever invented for that too."
His assessment is that AI probably has a negative impact on education today — cheating is rampant, and AI is blamed for lost math skills and failed classes — while noting the evidence cuts both ways: a randomized trial giving students unstructured ChatGPT access lowered retention-test scores, while a structured virtual-tutoring trial in Nigeria found substantial gains after six weeks using GPT-4. Teachers use AI for busywork, lesson plans, and feedback with mixed effects. He concludes that AI ultimately requires rethinking what students should learn. See AI in Education.
Data centers
Newman reports data centers as possibly generating more US headlines than AI's effect on jobs or education, citing more than 300 local and state bans or moratoriums. He separates the objections by merit: water-usage concerns are "in most cases significantly overstated," while grid-power scrambles and "temporary" gas and diesel generators, with their air and noise pollution, "have more validity." He notes some opposition is driven by concern about AI generally, argues bans seem likely to push construction into other jurisdictions and possibly outside the US, and states that done well, data centers can lower electricity rates, create jobs, and provide tax revenue. See AI Data Centers, Data Center Siting / AI Power Politics.
Sector survey
| Sector | Newman's assessment |
|---|---|
| General productivity | Thousands of small uses at work and home — "dumb little stuff" that adds up. A Pakistan study found judges using AI clear 6% more cases at slightly higher quality. |
| Software development | Large productivity boost available for coding, but results vary dramatically by tool use and work type; overall business impact generally much smaller, though some organizations are "cracking the code." |
| Health care | The biggest impact so far is individuals using chatbots for medical information — less useful than a doctor's focused attention, but often better than the actual alternatives. Also reduces clinicians' "pajama time" paperwork; doctors increasingly use tools such as OpenEvidence. |
| Medical science | A wide range of applications (AlphaFold, research agents), but the current wave of new treatments is mostly not AI-driven; the research-to-clinic pipeline takes decades. |
| Mental health | Deeply mixed — companion, coach, sycophant, crutch, or cheap substitute for human relationships; implicated in teen suicides and "AI psychosis." Used wisely it can help, but Newman suspects it mostly is not being used wisely. |
| Cybersecurity | Raises the temperature for attackers and defenders alike, with attacks already reported as carried out autonomously by AI. Newman leans negative in the near term and expects high-profile AI-enabled incidents. |
| War | Militaries process satellite imagery, select targets, and guide drones; AI support reportedly accelerated the US targeting cycle in Iran. |
| Politics | Matters as both an issue and a tool; as a tool it has had little visible impact, and predicted deepfake catastrophes have not borne out — probably, he suggests, because audiences are inoculated by human-authored fake news. AI-flooded public-comment channels are "a worrying sign." |
| Other | Substituting for web pages and how-to books, a boon for users and a hit to publishing incentives; sharply reducing the need for Be My Eyes human volunteers; enormous robotics investment with little to show yet. |
Conclusion
Newman's closing argument inverts the usual caution: "That absence of evidence is suggestive evidence of absence" — AI is, for the most part, not yet a primary factor driving most trends in society and the economy, "or at least it wasn't as of 6–12 months ago." He then argues 2026 may be the last year for which that holds, on the grounds that a broad range of AI metrics are growing at 3× to 10× per year and exponential growth builds rapidly.
His prescription is about measurement capacity rather than policy substance: invest in better and faster data collection and the organizational ability to respond quickly, closing with the New York Times' formulation that "by the time the data is clear, [economists] warn, it could be too late for policymakers to figure out how to respond."
Provenance
Retrieved from secondthoughts.ai, the author's own publication; the gap-identifier verified it on 2026-07-23 against the byline and a modified-time of 2026-07-20, and established that the canonical slug is /p/impact-of-ai. The raw file is a condensed rendering that preserves the essay's structure, figures, and quoted passages but summarizes some passages in the third person rather than reproducing them verbatim; direct quotations above are those the raw file marks as such.
Relationships
- supports: AI and Productivity — the empirical case that AI's macroeconomic impact is not yet separable from noise
- supports: AI Is Really Weird — reaches Kedrosky's conclusion by a broader route, adding sector-level survey to the macro-statistics argument
- contradicts (partial): claims of large near-term labor displacement — Newman bounds the Stanford "Canaries in the Coal Mine" result and argues no broad employment signal exists
- related: AI Labor Disruption, AI Data Centers, AI in Education, AI in Healthcare, AI Bubble vs. Buildout — Synthesis, AI Psychosis