Yampolskiy's 99% Unemployment Claim on DOAC: What Holds Up

The Diary Of A CEO interview with Roman Yampolskiy carries a blunt title: only five jobs will remain in 2030. The episode has drawn more than 23 million views since it aired in September 2025. One small problem rarely mentioned by its promoters: the list of five jobs is never read out during the 87-minute interview. What exists is a loose set of three categories with no counts behind them, all coming around the 35-minute mark: personal experience a machine cannot take over, the willingness of wealthy people to pay for human service, and handmade products [1]. Other predictions in the same video are far more specific, and far easier to test.

Verdict: right direction, numbers that leap

Method: the transcript was taken from the official channel, then every quantitative claim was tested against primary sources. The source list covers the OpenAI- and Oxford-authored research on job exposure, official pages of the US Department of Labor and the Bureau of Labor Statistics, Google DeepMind's official announcement, Alphabet's Waymo ridership data, the Metaculus forecasting aggregate, two YouGov surveys, and the reporting on OpenAI's Superalignment team being dissolved. Total: 11 sources, 11 word-for-word quotes. Status labels used: verified, verified with correction, contested, not verified.

The results split in two. The case that AI capability is climbing fast is strong. AI systems now reach gold-medal level at the International Mathematical Olympiad under official contest grading [4], and driverless taxis already complete more than 250,000 paid rides per week in the US [3]. What does not survive is the leap in the headline numbers: 99 percent unemployment by 2030, AGI by 2027, and the claim that 60 percent of today's jobs could be replaced by models that already exist. The summary is below.

Claim in the video	Status	Closest evidence
AGI arrives in 2027, per prediction markets	contested	Metaculus median for weakly general AI: 12 June 2028, range starting November 2026 [7]
99 percent unemployment within five years	not verified	Official research measures task exposure, not unemployment; BLS projects normal growth through 2034 [2][10]
60 percent of jobs replaceable by today's models	not verified	Closest verified figure: 19 percent of workers have at least half of their tasks exposed [2]
Humanoid robots replace plumbers by 2030	not verified	No official roadmap exists; BLS projects physical work growing at the average rate through 2034 [10]
OpenAI's Superalignment team dissolved six months after announcement	verified with correction	It was indeed dissolved; the correct gap is about one year (May 2023 to May 2024) [8]
From losing at math to Olympiad gold in three years	verified	Gemini Deep Think: official IMO 2025 gold, 35 of 42 points [4]

Testing the 99 percent unemployment figure

The heaviest claim in the video: the world will see unemployment levels never seen before, not 10 percent, but 99 percent [1]. Against primary sources, the number has no counterpart. The research most often cited in this field, written by OpenAI researchers with Daniel Rock, measures task exposure rather than unemployment: around 80 percent of the US workforce holds jobs with at least 10 percent of tasks exposed, and around 19 percent of workers hold jobs with at least half of their tasks exposed. The same paper explicitly declines to predict an adoption timeline [2]. Official BLS projections point the other way for physical work: the transportation and material moving group, one of the largest occupation groups in the US, is projected to grow about as fast as the average for all occupations through 2034 [10]. Frey and Osborne's reappraisal at Oxford, ten years after their famous 47 percent figure, underlines the deployment bottleneck: high-stakes settings keep humans in the loop because mistakes are costly or irreversible [9]. The 99 percent figure is a speaker's extrapolation, not a measurement.

The AGI-by-2027 prediction is claimed to rest on prediction markets and lab CEOs [1]. The prediction market is real, but the current Metaculus median for weakly general AI is 12 June 2028, with a range that begins in November 2026 [7]. Near dates do sit inside the distribution; they are simply not its center. The claim of "two to three years" picks the lower tail and presents it as consensus.

The claim that 60 percent of today's jobs could be replaced by existing models matched no source study [1]. The closest verified figure remains the 19 percent of workers with half of their tasks exposed [2]. On the meaningless-work premise, the data is more human: a 2015 YouGov survey found 37 percent of British workers saying their job made no meaningful contribution to the world, and the 2024 repeat survey recorded 33 percent [5][11]. That is worker perception rather than a measure of machine substitutability, and its stability at one third across nine years suggests the phenomenon is neither new nor waiting for AGI.

The promise of humanoid plumbers by 2030 is the easiest claim to test, because the deadline is close [1]. No official roadmap supports it, while official BLS projections still show normal growth for physical occupations [10]. This claim simply waits: in 2030, everyone can check it themselves.

One factual correction deserves a sentence of its own. Yampolskiy says OpenAI's Superalignment team promised to finish within four years and was dissolved six months later [1]. The official chronology: the team was announced in May 2023 with a commitment of 20 percent of computing capacity over four years, then dissolved in May 2024 after Ilya Sutskever and Jan Leike left [8]. The real gap is about one year, not six months. The substance of the claim still stands: a safety team mandated to control systems smarter than humans really was dissolved.

What remains unverified

Three other positions in the video are reported as the speaker's opinions rather than verified facts: the near-certainty that this world is a simulation, the belief that longevity is one breakthrough away, and the belief that Bitcoin is the only genuinely scarce asset [1]. None of the three has a near-term way of being tested, and none is counted in the verdict.

For readers, one pattern from this fact-check is reusable. Predictions that carry ranges and official measurements tend to hold: 19 percent of workers with half their tasks exposed, a 2028 median, IMO gold at 35 of 42 points [2][4][7]. Predictions that carry a single dramatic number with no study behind them, 99 percent and 60 percent, found no source [1]. What is verified from this interview is already serious enough without inflation: capability is rising faster than control, and the team mandated to bridge that gap was disbanded [4][8]. This article's judgment: Yampolskiy's strength lies in the evidence that has already happened, not in his expired dates.

Sources:

[1] DOAC: The AI Safety Expert - These Are The Only 5 Jobs That Will Remain In 2030 - Dr. Roman Yampolskiy (2025-09-04, 1h27m, 23,273,329 views as of 2026-10-05) [2] Eloundou, Manning, Mishkin, Rock: GPTs are GPTs - An Early Look at the Labor Market Impact Potential of LLMs (arXiv 2023, later in Science) [3] CNBC: Waymo reports 250,000 paid robotaxi rides per week in U.S. (2025-04-24) [4] Google DeepMind: Gemini Deep Think achieves gold-medal standard at IMO 2025 (2025-07-21) [5] YouGov: 37% of British workers think their jobs are meaningless (2015) [6] U.S. Department of Labor: History of Federal Minimum Wage ($7.25 federal floor since 2009) [7] Metaculus: When will the first weakly general AI system be publicly announced? (community estimate) [8] CNBC: OpenAI dissolves Superalignment AI safety team (2024-05-17) [9] University of Oxford: Frey & Osborne, Generative AI and the Future of Work - A Reappraisal [10] U.S. Bureau of Labor Statistics: Transportation and Material Moving Occupations (OOH) [11] YouGov: What are the most meaningless jobs? (2024 repeat survey, 5,889 workers)