SABINE HOSSENFELDER: Forecasts that AI would quickly automate jobs and physical work have run well ahead of reality

The Gist

Sabine Hossenfelder says experts got jobs wrong the other way. Truckers were supposed to be replaced early, but only a few hundred driverless trucks run today. The World Economic Forum expected machines to do half of work tasks by 2025, and it found about 22 percent done mainly by machines, plus 30 percent done jointly. So experts can be too fast as well as too slow. This packet steelmans the argument Sabine Hossenfelder makes in this video, in its best light, for analysis; it is not an endorsement of her conclusions.

Conclusion

Expert error about AI has not run in only one direction: forecasts that AI and automation would rapidly take over jobs and physical work, such as truck driving, have run ahead of what has happened, so the same experts who underestimated capability have overestimated deployment in some domains.

Premises

1. In 2020 the World Economic Forum forecast that by 2025 machines would perform as much of current work tasks as humans, and in 2025 it found that about 22 percent of tasks were done mainly by technology.
2. Truck driving was widely expected to be among the first jobs automated, yet the leading U.S. driverless trucking company expects to operate only a few hundred driverless trucks by the end of 2026, against about two million heavy truck drivers.
3. Expert panels have overestimated physical-world AI adoption, including autonomous ride-hailing and AI's help to novices on laboratory tasks.
4. Experts have therefore erred in both directions depending on the domain, not uniformly toward caution.

Assumptions

• The 2020 forecast concerned the share of time on current tasks done by humans versus machines, and the 2025 report found 47 percent of tasks done mainly by humans, 22 percent mainly by technology, and 30 percent by a combination, with a near even split now expected by 2030, so the forecast looks delayed rather than simply wrong.
• The forecasting group finds too little resolved evidence to judge expert forecasts of AI's labor and macroeconomic effects, and on some adoption measures, such as the share of U.S. work hours assisted by AI and AI company revenue, forecasters underestimated rather than overestimated.

Analysis

Argument type: Inductive.

Potential Fallacies

• Equivocation on 'the same experts' (Conclusion) — The conclusion implies one group made both errors. The premises draw on different communities: WEF survey respondents, industry and company forecasters, and forecasting panels. No premise shows these overlap with the people who underestimated capability. Errors in different directions across the wider expert ecosystem are shown, but not errors by an identifiable single group.
• Conflating capability with deployment (Conclusion and P4) — Underestimating what AI can do and overestimating how fast it is adopted are errors about different variables. Capability depends on research progress, while deployment depends on cost, regulation, liability and infrastructure. Both errors are compatible with one coherent view in which capability advances fast and diffusion is slow, so pairing them as 'errors in both directions' suggests a symmetry that may not exist.
• Hasty generalization / selective examples (P2, P3, P4) — Trucking, ride-hailing, wet-lab uplift and the WEF task share are vivid, memorable cases, not a systematic sample of forecasts. Without base rates, the number of overestimates and underestimates cannot be compared. The conclusion's hedge ('in some domains') and the A2 concession limit the damage, but the evidence remains illustrative.
• Circular support (P4 restates the conclusion) (P4 and Conclusion) — P4 is introduced with 'therefore' and largely restates the conclusion, so it adds no independent support. It is better read as an intermediate summary of P1–P3 plus A2.
• Timing error treated as directional error (P1, P2 with A1) — A forecast that arrives late but converges (as A1 suggests for the WEF case, with a near-even split by 2030) is an error about pace, not about where the trend is heading. Calling it an overestimate is accurate for the 2025 date but says little about long-run judgment, and the same reading may apply to trucking and ride-hailing.
• Mismatched benchmark (P2) — Comparing a few hundred trucks with two million drivers makes the shortfall look dramatic, but the relevant comparison is with what specific forecasters predicted for a specific date. Early-stage deployment on an S-curve is also not well measured against the whole workforce.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Capability and deployment are different forecasting targets made largely by different people. Underestimating capability while overestimating deployment is consistent with one coherent model of AI progress, so it is a false symmetry. The sign of error on the variable that drives risk and policy, capability, remains underestimation.
• Premise 1 (Medium impact) — By A1, the WEF forecast looks delayed rather than wrong, with a near-even split expected by 2030. This is a timing miss, not a failure of direction, and the 'combination' category suggests machine involvement is larger than 22 percent implies.
• Premise 2 (Medium impact) — Slow driverless trucking may reflect regulatory, safety-validation, liability and capital constraints, along with the long tail of edge cases, rather than expert misjudgment. Early deployment on an S-curve also makes the current count a weak indicator of the trajectory.
• Premise 3 (Medium impact) — Robotaxi services have scaled quickly in several cities, so 'overestimated' depends on the forecast year and baseline. Wet-lab uplift concerns capability in a narrow study, not adoption, and does not belong under the same heading.
• Premise 4 (Medium impact) — On the adoption measures most relevant to near-term economic impact (AI-assisted work hours, AI company revenue), forecasters underestimated. The net bias may therefore still lean toward underestimation, which would support the view the argument seeks to temper.
• Premise 1 (Medium impact) — WEF reports and company timelines are employer surveys or promotional statements, not calibrated expert forecasts. Their misses say little about the reliability of expert judgment.

Suggested Improvements

• Separate capability from deployment — State explicitly that the thesis is that capability was often underestimated and physical deployment often overestimated, and drop 'the same experts' unless person-level data show overlap. This removes the equivocation and makes the claim exactly what the evidence supports.
• Anchor forecasts to named sources and dates — Cite specific dated predictions (who forecast what, by when) for trucking and ride-hailing, and compare outcomes against those targets instead of total driver headcount. Makes the size and timing of the miss assessable and avoids a misleading scale comparison.
• Distinguish delay from error — Apply the delayed-not-wrong analysis from A1 consistently to trucking and ride-hailing, and report growth trajectories (fleet size, paid rides) alongside snapshots. Lets readers judge whether forecasts were mistimed or misdirected.
• Broaden the evidence base — Add a systematic set of resolved forecasts with calibration scores by domain and forecaster group, including cases of underestimation. Moves the argument from vivid anecdotes to evidence of whether error is systematic, such as digital capability versus physical deployment.
• Fold counter-evidence into the main argument — Move the underestimation evidence (AI-assisted work hours, AI company revenue) out of the assumptions and into the premises, and discuss what it implies for net bias. Strengthens the 'both directions' claim and pre-empts the objection that the evidence is one-sided.
• Guard against misuse — Add a caveat that two-sided error implies wide uncertainty, not that warnings about capability or displacement can be discounted. Prevents the conclusion being used to justify complacency or, conversely, to justify aggressive timelines.

Scenario Tests

• Driverless trucking scales sharply between 2026 and 2030, and a 2030 resurvey shows a near-even split of human and machine tasks. (Challenges) — The overestimates would be reclassified as timing delays, weakening the case that experts misjudged deployment, though the 2025–2026 shortfalls would remain real misses.
• A systematic database of dated AI forecasts shows capability underestimated in digital domains and deployment overestimated in physical domains. (Supports) — Supports a structured, domain-dependent two-directional pattern, with the caveat that it would favor a capability-versus-deployment reading over a claim of general expert unreliability.
• Person-level forecast data show that individuals who underestimated capability were not the ones who overestimated deployment. (Challenges) — The 'same experts' wording fails, although the weaker claim that errors go both ways across the ecosystem survives.
• Further resolved forecasts show AI-assisted work hours and revenue were underestimated across many adoption measures. (Challenges) — Net bias may lean toward underestimation, which would shift the emphasis away from overestimation of deployment.
• Regulatory and liability barriers, not technical limits, explain most of the trucking and robotaxi delay. (Neutral) — Deployment forecasts were still too fast, so the descriptive claim holds, but the lesson concerns forecasting institutional friction and not the reliability of AI expertise.

Coherence & Relevance

The argument is coherent as a modest inductive case that forecast error has run in both directions across domains, and the stated assumptions candidly supply the main qualifications: delay, not failure, in the WEF case, and underestimation on some adoption measures. Coherence weakens where the conclusion reaches for 'the same experts' and pairs capability underestimation with deployment overestimation as if they were one forecasting task. The headline 'run well ahead of reality' is also somewhat stronger than the delay-based evidence warrants. Restated as 'errors have gone both ways across domains, with physical deployment often slower than forecast', the argument is reasonably well supported.

• In 2020 the World Economic Forum forecast that by 2025 machines would perform as much of current work tasks as humans, and in 2025 it found that about 22 percent of tasks were done mainly by technology. (Strong) — Directly illustrates overestimated near-term automation, but it is a survey-based forecast, not an expert one, and A1 frames it as delay. It does not speak to capability underestimation.
• Truck driving was widely expected to be among the first jobs automated, yet the leading U.S. driverless trucking company expects to operate only a few hundred driverless trucks by the end of 2026, against about two million heavy truck drivers. (Strong) — Concrete, high-profile example of the title's claim, but the forecast target and date are unspecified and the headcount comparison is not a forecast-versus-outcome comparison.
• Expert panels have overestimated physical-world AI adoption, including autonomous ride-hailing and AI's help to novices on laboratory tasks. (Moderate) — Broadens the evidence beyond trucking but groups different constructs (adoption vs. capability uplift) and is partly correlated with the trucking evidence.
• Experts have therefore erred in both directions depending on the domain, not uniformly toward caution. (Moderate) — Follows from P1–P3 plus A2 for the existential claim, but the underestimation half rests on A2 and not on the premises, and the step to 'the same experts' is not supported.

View this argument on LogicFirst.ai