Peter Thiel famously quipped, "We wanted flying cars, instead we got 140 characters” and the question "Where is my flying car?" has become a symbol of unfulfilled futuristic expectations. Personally, I never grew up with the Jetsons and the expectation of flying cars. However, I certainly anticipated the arrival of self-driving cars. In my early 20s, learning to drive felt as antiquated as learning to ride a horse—an outdated skill that seemed destined for obsolescence. But now, having reached my early 30s, I must reluctantly admit defeat and learn to drive. So, the question, "Dude, where is my self-driving car?" resonates with me.

Yet, as a non-driver living in Switzerland where trains are ubiquitous, a second, even more perplexing thought has recently struck me: "Dude, where is my self-driving train?". 

Both examples reveal that actual automation can significantly lag behind the techno-economic frontier of what could be automated.

The puzzle of self-driving trains

Automation has proven particularly effective in environments where there is a lot of data, a clear objective, a defined application scope, and relatively static conditions. Trains and metros operate on fixed routes at predetermined speed levels, something close to an ideal scenario for automation. Further, modern cameras and object recognition can offer reliable monitoring to prevent collisions with other trains or objects on the rails. Lastly, remote human monitoring and intervention from control centres can provide an additional layer of security. 

The International Association of Public Transport has standardized levels of train automation. The technology for the highest level of automation -  “Grade of Automation Level 4” or unattended train operations - has existed for about three decades. Examples include the Tokyo Yurikamome Line (1995); the Paris Metro Lines 14 (1998), 1 (2012) & 4 (2022); the Barcelona Metro line 9 (2009); the Milan Metro Lines 5 (2013) & 4 (2022); or the Rome Metro Line C (2014). 

Moreover, if we look at some ballpark numbers, it seems like automating trains and metros should be economically attractive. In the United States, operations expenses (labor & fuel) generally represent about two-thirds of the overall costs of transit. For example, for the MTA New York City Transit labor represents about 60% of total expenses and about 20% of employees are involved in subway operations.

Yet, surprisingly, with a few exceptions, such as airport movers, Copenhagen (2002), Singapore (2003), and Dubai (2009), most of the train and metro systems are not automated. Most of the metros of New York, Tokyo, London, and Paris are still operated by a human driver. Even in rich countries like Japan, Switzerland, and France nearly all trains remain human operated.

The puzzle of self-driving cars

Building self-driving cars is inherently more challenging than building self-driving trains. The road environment is a much more dynamic landscape filled with much more obstacles, such as pedestrians, cyclists, and other vehicles. So, in some ways, the technology for autonomous driving has come a long way. However, compared to the initial expectations, it is surprising how slow the roll out of autonomous cars has been: 

• Elon Musk predicted Tesla cars would be capable of full autonomy by 2017.
• GM announced it would mass-produce cars without steering wheels by 2019
• Daimler wanted to develop fully autonomous vehicles by the early 2020s
• BMW wanted to bring fully autonomous vehicles to the market by 2021.

Yet, in 2024, fully autonomous cars remain a rarity and are not available for individual purchase. Waymo operates 300 robotaxis in San Francisco with plans to expand further. Still, there is not much there yet beyond a few geographically limited trials. According to the Bureau of Labor Statistics the US economy alone still employs about 400’000 taxi drivers, 500’000 bus drivers, 1.7 million delivery truck drivers, and 2.2 million heavy truck drivers. 

What’s even more surprising: According to Swiss Re, one of the world's leading reinsurance companies, autonomous cars already outperform human drivers in terms of real-world safety. In over 3.8 million miles driven without a human being behind the steering wheel, the autonomous driving company Waymo incurred zero bodily injury claims in comparison with the human driver baseline of 1.11 claims per million miles. The Waymo Driver also significantly reduced property damage claims to 0.78 claims per millions miles in comparison with the human driver baseline of 3.26 claims per millions miles. So, it seems fair to say that the state-of-the-art of level 4 automated driving is already 3-4 times safer than human driving.

Why is automation so delayed?

The delayed adoption of automation is not unique to the transport industry. According to an analysis by McKinsey in 2017: “Almost half the activities people are paid almost $16 trillion in wages to do in the global economy have the potential to be automated by adapting currently demonstrated technology (...) While less than 5 percent of all occupations can be automated entirely using demonstrated technologies, about 60 percent of all occupations have at least 30 percent of constituent activities that could be automated.” 

The reasons for the adoption delays are diverse, still, the following are three important factors:

Limited economic competition: Trains and metros are text-book examples of natural monopolies with high entry barriers due to substantial upfront investment in tracks, stations, and rolling stock. Economies of scale and physical space requirements put further limits on the possibility of parallel systems. This means that there is limited economic pressure and there is simply no competitor that can offer New York Metro tickets at a 10% discount because they reduced operating expenses with automation. Economic profit is also often not the only goal of public transport companies (e.g., maintaining service levels to satisfy political constituencies).

Social resistance: Employees that fear job loss will often resist automation. This can include strikes and political lobbying by unions, such as the Teamsters’ opposition to local permissions to operate autonomous vehicles. There are also more hidden ways to drag the process out. For example, the Paris Métro had a plan for full automation as early as 1986. However, based on Capital magazine, the management later decided that it’s in its interest to only advance at a snail’s pace. This included “an unwritten agreement with the unions for thirty years not to do more than one automated line at a time" and the artificial inflation of automation costs by combining it with other items such as new trains or redoing the tiles at stations.

Legal barriers: Laws slowing down automation are largely downstream from social resistance. First, laws can outright ban or disincentivize automation. For example, in the US, the Urban Mass Transportation Act of 1964 required local agencies accepting federal transit grants to protect ”individual employees against a worsening of their positions with respect to their employment”. This provision has remained until today, known as Section 13(c). According to a 1976 (!) report on train automation by the Office of Technology Assessment, it “allows the elimination of jobs, but only as workers presently holding those jobs retire or vacate the positions for other reasons.” 

Second, laws may hold automated processes to more rigorous safety standards than human operators or as Michael Woudenberg calls it “Zero Forgiveness for Technology”. Imagine for a moment that we treated autonomous driving systems like human drivers. To obtain a national driving license that is internationally recognized an autonomous driving system would need to pass:

• A written test of knowledge of traffic laws. These are often multiple choice and you don’t need a perfect grade to pass. Autonomous driving systems are programmed with extensive knowledge of traffic laws and should be able to pass this with flying colors. 
• A practical driving test at any municipality of choice. Autonomous vehicles can perform complex manoeuvres and navigate through urban, suburban, and highway settings. Again, it seems that the leading autonomous driving systems should be able to pass the vast majority of driving tests.

In short, not every market has strong competitive pressures to automate, there can be social resistance to automation, and sometimes legal requirements slow down automation. So, even if Sam Altman would declare tomorrow that he has built a digital God and that he will roll it out for free to everyone, this would not immediately lead to full automation.