AI Futures

— where is artificial intelligence actually heading?

This page is written for a Toronto reader with no background in AI. It is not a hype piece and not a doom piece. It reports what named researchers, surveys, and institutions actually say about where AI is heading — including how much they disagree with each other — with a date attached to every number.

What’s actually happening

Two credible sources give very different answers to a simple question — “how many businesses use AI?” — and that gap is itself useful to understand. Stanford’s 2026 AI Index (published April 13, 2026) found 88% of surveyed organizations used AI in at least one business function in 2025, up from 78% in 2024. The U.S. Census Bureau’s own randomized national survey, covering December 2025 to May 2026, found only 17–20% of all U.S. businesses use AI, rising to 37% for firms with 250 or more employees. Both are real; Stanford’s survey skews toward larger, tech-forward companies, while Census captures the long tail of small firms that barely use it at all.

The cost of running AI models has fallen fast: Epoch AI, a research group that tracks this as measurement rather than prediction, finds the price of matching a given capability has dropped by roughly 50 times per year on average (data through March 2025), and by roughly 200 times per year looking only at data from January 2024 onward. Training compute for the largest models has grown roughly 5 times per year since 2020 (Epoch AI, tracked through February 2026).

Productivity gains are real but task-specific, not economy-wide. Named studies find: customer-support agents resolved 14–15% more issues per hour using an AI assistant (Brynjolfsson et al., 2025); software developers using GitHub Copilot completed 26% more pull requests (Cui et al., 2025); marketing teams using AI for ad creation saw a 50% increase in output per worker (Ju and Aral, 2025); and a broader study of European businesses found roughly a 4% labour-productivity boost from AI adoption (Aldasoro et al., 2026).

One narrow, replicated signal exists in the labour market: U.S. software-developer employment for workers aged 22–25 fell nearly 20% since 2024. Two young-people surveys, fielded independently, show the same underlying pattern: the Harvard Youth Poll (fielded November 3–7, 2025; 2,040 Americans aged 18–29) found 59% see AI as a threat to their own job prospects. Gallup’s Gen Z survey (fielded February 24–March 4, 2026) found weekly AI use holding steady at about 51%, while reported excitement about AI fell from 36% to 22% and hopefulness fell from 27% to 18% over one year — even among daily users.

Where credible people think it’s heading

Metaculus, a public forecasting platform, runs a long-tracked question on when the “first general AI system” will be built. As of the 2026-08-05 re-check, the community median sits at December 2032, drawn from roughly 1,900 forecasters, with a 90% range running from May 2029 to March 2040. That number has moved earlier and later repeatedly over the past few years, and what counts as “general AI” for the question to resolve is itself disputed — a live methodological problem, not a settled one.

The same underlying question, asked two different ways, produced wildly different answers from the same pool of people. A 2023 survey of 2,778 published AI researchers found a 50%-probability date of 2047 for AI that could do every task better and more cheaply than a human worker. But when a random half of the same respondent pool was instead asked about “full automation of labor” — a related but differently worded question — the 50%-probability date came back as 2116, nearly 70 years later. Same survey, same people, a 69-year gap driven only by how the question was phrased.

On catastrophic risk, the same 2023 survey found the median researcher put a 5% probability on AI causing human extinction or similarly severe outcomes, with a much higher average of 16.2% (roughly the odds of one round of Russian roulette). A separate structured tournament in 2022 paid 80 subject-matter experts and 89 “superforecasters” (people with strong, tracked forecasting records) to debate the question for months. Experts’ median estimate of AI-caused extinction by 2100 was about 3%; superforecasters’ median was about 0.38% — a ten-times gap. On most other risks in the same tournament, the two groups moved closer together after debating each other’s evidence. On AI risk specifically, they did not move at all, even after what the researchers describe as millions of words and thousands of forecasts exchanged.

What’s honestly uncertain

Two detailed documents about AI’s near future circulate widely and need to be read as what they actually are: scenarios, not forecasts.

“AI 2027” (Daniel Kokotajlo, Eli Lifland, Thomas Larsen, and Romeo Dean; published April 3, 2025) is a detailed, month-by-month story of a possible AI “intelligence explosion.” The authors themselves call it a scenario, not a prediction, and say their own collective range for the events depicted spans roughly 2028 to 2032.

“Situational Awareness: The Decade Ahead” (Leopold Aschenbrenner, published June 2024) argues AI systems capable of doing their own AI research could arrive by 2027. A financial disclosure matters here: after the essay went viral, Aschenbrenner founded an investment fund built around the essay’s own thesis. As of a June 2026 report, that fund held more than $20 billion in assets. By late July 2026, following a forced sale of leveraged positions, assets reportedly fell to roughly $10 billion. The essay’s author has a direct, disclosed financial stake in the AI-acceleration case the essay argues for — a fact that should travel with the essay every time it’s cited.

What counts as “AGI” or “general AI” at all is itself a live dispute among the people building these forecasting tools, not just a public misunderstanding.

Why it matters for Toronto

The source file behind this page does not localize its figures to Toronto specifically. What it does show is that the entry-level hiring effect for workers aged 22–25 in AI-exposed roles is a general pattern, not a Toronto-specific one — and it would apply here too. For Toronto’s own labour-market numbers, including the city’s current unemployment rate, see the Work & Economy page in this library.