Scoring 2020's self-driving predictions, five years on

In the late 2010s, “full self-driving next year” was an annual tradition. A cohort of those calls had hard deadlines around 2020. Years on they can be graded, and the misses are as informative as the hits.

The calls

These were not fringe claims. They came from the companies best placed to know, said in public, with dates attached.

In 2016 Ford announced it would have a fully autonomous vehicle, with no steering wheel and no pedals, in commercial ride-hailing service in 2021. The same year Lyft’s co-founder John Zimmer wrote that a majority of Lyft rides would be in autonomous vehicles within five years. GM’s Cruise set out plans to launch a commercial robotaxi service in 2019. And at Tesla’s Autonomy Day in April 2019, Elon Musk said the company expected more than a million robotaxis on the road in 2020.

Each of those is specific enough to grade. That is rarer than it should be, and it is the first thing to look for in any prediction: could you tell, on the day, whether it came true?

What happened

None of them came true on time, and some of them never will.

Ford and Volkswagen shut down Argo AI, the self-driving company they had funded, in 2022. Cruise’s 2019 launch slipped, and in October 2023 California suspended its driverless permits after one of its cars dragged a pedestrian who had been knocked into its path. GM stopped funding Cruise’s robotaxi work in December 2024. Lyft’s rides stayed overwhelmingly human-driven. Tesla’s million robotaxis did not appear in 2020.

The nearest thing to a hit is Waymo, which opened fully driverless rides to the public in the Phoenix area in October 2020, and has since expanded paid driverless service to a handful of US cities. That is a real achievement and a slow one: city by city, geofenced, years after the first demos. It looks nothing like the overnight switch the 2020 calls described.

What the misses have in common

The failed predictions were not random. They shared a pattern: a dramatic demo, a straight-line extrapolation from it, and a confident date attached to a long-tail problem. The demo was real; the extrapolation was the error.

Driving is almost entirely long tail. A system that handles 99% of situations well has done the easy part. The remaining fraction is made of things that are each individually rare and collectively constant: a road crew waving traffic through a red light, a plastic bag that looks like a child, a pedestrian thrown into the road by another car. Every demo shows the 99%. The deadline is set by the rest.

There was a second shared mistake, quieter than the first. Most of the calls treated the technology as the only hard part. Regulators, insurers, city governments and public trust all had to move too, and none of them moves on an engineering roadmap. The Cruise suspension was a regulatory event triggered by a single incident, and it ended a programme that had cost billions.

How the scoreboard grades a call

Calls like these go on the scoreboard after the fact, as resolved history, and are marked that way. The board lists “fully autonomous robotaxis available to the public nationwide by 2020” as a miss.

It does not only list misses. The same board records that AI would beat a top professional at Go, settled by AlphaGo in 2016. It records that a model would pass the bar exam at the level of a competent attorney, settled by GPT-4 in 2023, and that an AI would win gold at the International Mathematical Olympiad, settled in 2025 by systems from both Google DeepMind and OpenAI. Calls that came true are marked as plainly as the ones that did not.

The open AGI calls beside them are the ones that test calibration: each is listed with who made it and the year it falls due, stays on the board until that year arrives, and is then graded in public, misses included. That includes Ray Kurzweil’s long-standing 2029, Shane Legg’s even odds by 2028, and a community forecast that an AI agent will complete a multi-day software project on its own by 2027. The point is not to dunk; it is to calibrate.

Why self-driving has no date on this site

The site tracks self-driving as one of its ten domains, and it publishes no forecast for it. The only public series is a community tracker of miles between driver interventions, built from owners’ own telemetry. That is useful, and it is not a controlled benchmark: the drivers, routes and software versions all change underneath it.

The figure reported from it is about 0.6 doublings a year, roughly five times slower than the software domain. If that is anywhere near right, the gap between how fast AI is improving at writing code and how fast it is improving at driving a car is itself the lesson of 2020. Progress in one domain says little about another, however much the demos look alike.

The lesson carries straight into AGI timelines. When a striking capability demo lands this year, the question is which part is the demo and which part is the deadline.