Independent New Hampshire Journalism · Northfield, NH

When Will AI Surpass Humans? The Builders Just Moved Up the Date

The people selling artificial intelligence say a decade or less. The largest survey of AI researchers says 2047 for every task, and that guess jumped 13 years closer in a single year. Here is the timeline, with receipts.

The people building artificial intelligence now say, in public and on the record, that machines will out-think humans at most paid work within a decade. That is the claim. The evidence behind it is messier, and more interesting, than either the hype or the eye-rolling lets on.

This piece does one job: lay out when machines are likely to pass us, milestone by milestone, using only sources you can open and check. Nobody knows the date. But the direction is not in dispute, the speed is now measurable, and every serious forecast keeps moving the same way. Closer.

What the builders are promising

OpenAI chief executive Sam Altman opened January 2025 by writing on his blog that his company is now “confident we know how to build AGI” and predicting that AI agents would join the workforce that same year. Anthropic’s Dario Amodei, in his October 2024 essay Machines of Loving Grace, wrote that powerful AI could arrive as early as 2026, while allowing it could take much longer. His definition deserves a second read: a system smarter than a Nobel Prize winner across most fields, running as millions of copies at many times human speed. He calls it “a country of geniuses in a datacenter.”

Google DeepMind’s Demis Hassabis, the most careful of the three, told reporters at a March 2025 London briefing, as reported by CNBC, that AI matching humans at any task will start to emerge within five to ten years. He has repeated the window since, comparing the coming shift to the Industrial Revolution, only larger and compressed into a decade instead of a century.

One caution before you bank any of that: all three men run companies whose valuations depend on the claim being true, and each is raising historic sums on it. Their word is a data point, not a verdict. So check the scoreboard instead.

What the machines have already taken

The list of things computers do better than the best living human stopped being a novelty act years ago. Chess fell in 1997, when IBM’s Deep Blue beat world champion Garry Kasparov. Go fell in 2016 to DeepMind’s AlphaGo. Then the wins left the game board. DeepMind’s AlphaFold cracked the 50-year problem of predicting protein structures and mapped roughly 200 million of them, work the Royal Swedish Academy of Sciences honored with the 2024 Nobel Prize in Chemistry for Hassabis and his colleague John Jumper.

In July 2025 the line moved again. A Google DeepMind model earned an officially graded gold medal at the International Mathematical Olympiad, scoring 35 of 42 points and solving five of six problems inside the same four-and-a-half-hour sessions the students faced. Roughly 8 percent of the world’s best young mathematicians reach gold in a given year. The machine did it in plain English, no special notation required.

Notice the pattern. Every time a domain falls, we quietly redefine real intelligence as whatever is left.

The trend you can measure

The most useful number in this debate comes from METR, an independent evaluation nonprofit. Its researchers measure the length of tasks, timed against skilled humans, that frontier AI systems can finish on their own at least half the time. That horizon has doubled roughly every seven months for six straight years. When the study landed in March 2025, models were near-perfect on tasks that take a person under four minutes and failed most tasks past four hours. Run the curve to the end of the decade and autonomous agents are completing month-long projects.

The same research explains why today’s systems feel both magical and useless. They beat experts on most exam-style tests while flunking work a competent office temp handles without thinking, which is why the people who study this talk about a jagged frontier: superhuman on some peaks, stranded in the valleys. The peaks keep multiplying.

The brakes are real too, and the boosters admit it. Amodei’s own essay lists the limits intelligence cannot think its way past: experiments that run at the speed of cells, hardware that ships at the speed of factories, and laws that change at the speed of legislatures. Smarter is not the same as instant.

What researchers say when nobody is selling

The largest survey of the field, published in the Journal of Artificial Intelligence Research, asked 2,778 researchers who publish at top AI venues to put dates on the finish line. Their aggregate answer: a 10 percent chance that unaided machines outperform humans at every possible task by 2027, and a 50 percent chance by 2047. That 2047 figure arrived 13 years earlier than the same survey produced only one year before.

Ask the same people when every human occupation becomes fully automatable, the true “all things” bar, and the median answer retreats to 2116. Yet even that century-scale guess fell 48 years in a single survey cycle. A forecast that moves 48 years overnight is a shrug with error bars. Both shrugs moved the same direction.

In one year, the field’s median forecasts for both milestones moved only one way. Chart: Granite State Report. Data: Grace et al., Journal of Artificial Intelligence Research (2025).

The survey holds one more number that would end careers in any other industry. Depending on how the question was phrased, between 38 and 51 percent of those researchers put at least a 10 percent chance on advanced AI producing outcomes as bad as human extinction. The field’s own researchers said that about the technology they work on.

The skeptics deserve their paragraph, and they carry numbers too. In a 2025 study by the Association for the Advancement of Artificial Intelligence, 76 percent of 475 survey respondents called it unlikely or very unlikely that simply scaling today’s approaches produces general intelligence. They may be right. This field has broken its promises before, twice, badly. But listen to what the argument is now about. The mainstream dispute concerns the route and the decade, no longer the century.

The betting markets split the difference. An aggregation of forecasts (AGI Timelines Dashboard, Goodheart Labs) from Metaculus, Manifold, and Kalshi put the median arrival of AGI at 2031 as of July 19, 2026, with an 80 percent confidence band running 2027 to 2044. Wide. And still inside many readers’ working lives.

The timeline, milestone by milestone

Already here. Superhuman at chess, Go, protein-structure prediction, Olympiad mathematics, and most exam-style knowledge tests. Not yet trusted to run your errands.

2026 through 2028: the agent years. This is the window the labs themselves named. Amodei said powerful AI could land as early as 2026. The researcher survey gives 10 percent odds on everything by 2027. METR’s curve points to systems handling days-long assignments without supervision. Amodei also told Axios in May 2025 that AI could erase half of all entry-level white-collar jobs and push unemployment to 10 to 20 percent within one to five years; plenty of economists called that overblown, and Axios said so in its own reporting. Either way, the squeeze shows up as hiring freezes long before it shows up as pink slips.

2029 through 2035: the crowd’s window. The forecast markets’ median sits at 2031. Hassabis’s five-to-ten-year clock, started in 2025, runs out here. If METR’s doubling holds, agents finish month-long projects. This is the band where machines doing most cognitive work cheaper than people stops being a prediction and starts being a line item.

Around 2047: the researchers’ coin flip. Even odds, by the field’s own aggregate, that machines outperform humans at every possible task, better and cheaper. A reader who is 24 today would be 45.

2116, or never: everything, including the hands. Full automation of every occupation, the plumber and the hospice nurse included, sits a century out by the median researcher guess. Hold both numbers at once. The same experts say every task by 2047 and every job by 2116, those answers strain against each other, and the survey’s own write-up flags the tension. The contradiction tells you that nobody, including the people best positioned to know, has a working model of the endgame. What they agree on is the direction.

What this means for New Hampshire

Start with a claim New Hampshire holds on this story that predates the industry: the field got its name here. The 1956 summer workshop at Dartmouth College in Hanover coined the term artificial intelligence. Seventy years on, the question has come home.

New Hampshire is the advanced-economy case in miniature: a services state built on health care, education, professional work, and small business, and among the oldest states in the country by median age. The International Monetary Fund estimates about 40 percent of jobs worldwide are exposed to AI, rising to roughly 60 percent in advanced economies, with about half of those exposed jobs facing downward pressure on wages or existence. The state’s own numbers lean into the blast radius: NH Employment Security projections compiled in GSR’s 2025 workforce report show computer and mathematical occupations as the fastest-growing group through 2032, up about 18 percent. Those are the first jobs the labs learned to automate.

The Granite State has already met the sharp edge of this technology. Before the January 2024 presidential primary, an AI-generated robocall imitating President Joe Biden told New Hampshire voters to stay home. The legislature answered within months.

Concord regulated the symptoms, and quickly, to its credit. But the AI laws it passed govern how the technology gets used; none of them touches what it does to the labor market. Meanwhile the state still has not settled how it funds the schools that are supposed to prepare kids for whatever work remains; GSR’s case for stronger school funding reform lays out the stakes. The target those schools are aiming at moves every seven months.

The people betting your future will not live in it

Here is the part I keep coming back to. The dates above were set by executives spending other people’s billions, and the policy response, such as it exists, belongs to office-holders who will be retired, or gone, before the 2047 coin flip lands. I wrote a book about that pattern, Generational Malpractice: the people holding the wheel keep pricing a future they will never occupy, and the bill lands on whoever is young enough to still be here when it comes due.

So, the answer to the question in the headline. Will AI surpass humans in all things? By the field’s own median guess, not in your lifetime. At the things that pay New Hampshire’s mortgages, quite possibly within the decade. Nobody can hand you the year, and anyone who does is selling something. But when every serious forecast moves in only one direction, the question stops being whether. It becomes what we do with the years while the machines still cannot.

— Dexter Dow, Granite State Report

Fact check

Editor’s note. Every load-bearing claim above was checked against the listed source during production on July 20, 2026; every hyperlink in this article was opened and confirmed live on that date. CNBC, Axios and the Nobel announcement are cited as plain text. All dates for future events are forecasts attributed to the people and surveys that made them, not statements of fact. Corrections: Granite State Report corrects verified errors promptly and appends a dated note identifying what changed and when. The corrections policy and the corrections register are public, and correction notes are never removed.