The East India Company ship Warley was converted into the Royal Navy HMS Calcutta (left) to assist in armed transport. The French Régulus (right) also stranded on the shoals of Les Palles in 1809. Wikimedia Commons.

What happens to insurance once our cars are autonomous?

On April 25, 2017, I testified in Sacramento on upcoming Autonomous Vehicle regulation. Below is the letter addressed to the CA DMV.

In the spirit of John Brooks’ 1969 book Business Adventures, a favorite of both Bill Gates and Warren Buffet, where classic tales of Xerox, PARC and the IRS are shared, we draw from the early days of shipping to help illustrate what is going on in the world of AVs.
The early days of autonomous vehicles are very similar to the early days of clipper shipping. The designs of ships were highly secretive. Trade routes were hotly contested.

The know-how to build faster sails, more robust hulls, and shipyards (the predecessor to railroad organized labor and Model T assembly line) were highly guarded secrets. Every aspect of shipping was very much the zeitgeist of that era.

This zeal was coupled with the permissible possibilities that governments fought over in wars and reached settlement with treaties. Nations were economically incentivized to build navies to enforce these treaties and guarantee monopoly of trade to new colonies, especially with the lure of continued riches in the New World.

Today, the stakes are just as high. Just as ships of the era past needed to keep sailing at all times to map the world, the modern day autonomous vehicle must keep moving at all times to reach its fullest potential.

As of this writing, intellectual property is being litigated and AV strategy is reflected in private company valuation as well as public company market cap.

What is permissibly possible falls into what is addressed and not addressed in regulation.Enforcement falls to society at large, in part by regulators who set ground rules.

Enforcement will fall to the California Highway Patrol and various law enforcement agencies, by Courts who resolve disputes and clarify legal interpretations, by engineers who must run safety test, and by consumers who are the ultimate end-users.

But those who have skin in the game worry most about risk management: companies whose financial upside is tied to commercial success and the financial responsibility parties (e.g. lenders, insurers, or investors) who back the engineering process.

In the races of the past, however, I would argue that it was two non-engineering inventions that helped propel British companies to overtake their French and Spanish counterparts: 1) the joint-stock company in terms of financial innovation and 2) the risk syndicate in terms of insurance innovation.

Today, we see the same pressures, and new forms of financing and insurance need to work to truly spur the autonomous industry. This along with proper policymaking allows the founder engineer / corporate divisions to distribute reward and pool risk.

But while all human progress requires a dose of risk taking, it was through many cycles of instant wealth and financial ruin that yielded measured government response.

Quick examples include the failure of the British South Sea Company, which led to the Bubble Act of 1720 forbidding speculative formation of joint-stock companies unless approved with a royal charter. Similarly, the formalization of contracts to cover financial loss and human deaths later led to the Society of Lloyd’s registration at the Royal Exchange in 1774.

While we are in the early stages of learning about the potential benefits and drawbacks of AVs, we hope our policymakers today can draw lessons and wisdom of what British government was able to do and all other governments failed to do.

Given the complexity of the situation at hand, we appreciate the sizable responsibility that California Transportation Agency Secretary Brian P. Kelly and DMV Director Jean Shiomoto and their staff have in taking initiative to following Senate Bill 1298 Vehicle Code section 38750.

Possible Outcomes

Pontiac radiator shells, ca 1937. Pontiac was a brand of automobile manufactured and sold by General Motors between 1926 and 2010.

Due to this paradigm shift, many economic interests today are at stake, including those of banks, insurance companies, auto manufacturers, and new challengers.

While every player will have its own vision of the future unfolding, many investors have pontificated on the winning strategy and many founders have pitched their vision.

In parallel, several company board rooms and market pundits have had their own say, all of which make it ever more difficult for policymakers to digest.

Here are three major thematic paths that have been distilled. They are not mutually exclusive:

1) Software Trojan Horse

An existing manufacturer may look to deploy hardware technology needed for autonomous vehicle but simply “turn off” functionality in previous and future hardware builds.

This strategy allows human-aided cruise control or steering (Level 2) to essentially skirt DMV regulations and use that time to train software to maturity. When both regulation and software are ready, a simple software update will allow full Level 4 autonomy across a massive fleet.

Until then, the technology will most likely toggle between Levels 2 and 3, while insurance and litigation will focus on tested HCI or proper driver engagement.

2) Cash Rich

Large companies that have a substantial capital through a) revenue generated through a separate business line, b) private investor equity financing, or c) public equity or debt financing. They see their existing cash reserves and ability to self-insure as a competitive advantage, but this is short lived.

Ultimately, R&D budgets have an upper limit around the ~$10B mark, and a lot lower if they do find a renewable source to fund that R&D budget and transform it into generating revenue.

An alternative return on investment is the pressure to commercialize the R&D and generate interim revenue streams from robotaxi or fleet platoon operation. Ultimately a losing proposition for anyone is R&D not finding market success.

3) AV Medallion

Commercial fleet operators, those who already manage large fleets of ambulances, buses, trucks, taxis, and limos, evolve and begin to become certified AV operators within their respective specialty.

These existing fleet operators already use financing and insurance to cover their needs, and have existing end-customers with defined needs.

The Cash Rich Companies become Cash Poor after all the R&D and realize they would be better off in the business of selling or leasing autonomous vehicle to the commercial fleet operator.

Policymakers realize that piloting these vehicles require training, licensing, or medallion coordination.

The common denominator of the Path 2 and 3 is the assumption that commercial adoption will arrive before consumer adoption.

These fleet operators will pay up for the safety training, insurance, financing, and licensing just to attract the first wave of consumer adoption and economic efficiencies.

This momentum can be strong enough that consumers never have any interest in purchasing their own personal auto, in the same way consumers do not prefer to purchase their own ambulance for example.

This new class of vehicles may cost 5–10x as much as a normal human-aided vehicle, and would much rather defer the cost and expertise to a trained professional.

Alternatively, we can all easily see Path 1 arriving where one day that the technology costs get driven down and millennial consumers never buy cars at all but simply lease them and get a new one the following year.

This pattern already shows with the convenience of the Apple’s iPhone Upgrade Program, by which financing and warranty and contracts are completed with a few taps, and you get a new and improved iPhone every year.

Investor Pundits

Some investors, including Warren Buffet, have stated publicly that the arrival of self-driving cars will have a negative impact on Geico. Others have visions of autonomous vehicle replacing public bus systems.

Nevertheless, end of the day, consumers and businesses will find the path of least resistance whether it is a single or combination of paths outlined above. Pundits will have their say.

Nonetheless, the potential benefits are profound: local municipalities can replace empty buses with more efficient shuttles, robotaxi terminals next to every train station can increase ridership even more, and greater convenience and falling prices may finally alleviate freeway gridlock commonplace in California. In short, AVs, no matter who becomes the owner, becomes a public good.

In short, AVs, no matter who becomes the owner, become a public good.

Consumer demand for AV technology may put California on top, even against the most efficient transport infrastructure in cities like Singapore and heavily invested mega-regions such as Jing-Jin-Ji.

While a city-state may be more flexible in creating specific economic subsidies, and a central party government may have more authority pouring hundreds of billions of dollars towards a singular goal, California as a state has and continues to thrive with its creative and experimental nature.

We must be mindful though that our actions can have impact on the remaining 50 states and the larger federal system.

While CalSTA and DMV must stay impartial to singular commercial interest, it is helpful to think ahead of possible results from regulations that get implemented later in 2017 Q4 or 2018 Q1.

We would not be surprised if business and consumers change paths if they find a road block and switch tactics in their pursuit to match consumer demand to fuel R&D.

Trolley Problem Meets Moot Court

Alan M Turing with his colleague in 1951. Getty Images.

A popular philosophical exercise is to ask how artificial intelligence should decide between saving the lives of the vehicle’s passengers versus saving the lives of non-passengers.

Another popular exercise for legal scholars is to think through how an algorithm can be liable for negligence. However, without accidents, plaintiff attorneys, and actual case law, it will be many years before the community finds answers.

Until then, it may be productive to explain that engineering buzz words around “deep learning” or “neural networks” are simply used to describe mathematical strategies to solve problems.

The automata of these math problems at a massive scale leads to machine learning. The “learning” that gets done is based on a vast set of data collected. These techniques behind autonomous vehicle must be trained before they can get any good.

Moreover, many neural network or deep learning algorithms are often NP-complete problems, meaning that time needed to solve these algorithms increases quite quickly as the size of the problem grows.

Even with a supercomputer with the most advanced GPUs on board, some decisions cannot be solved with discrete certainty. From a legal scholar perspective, this throws a wrench into the trolley problem because decisional liability is pitted reasonable person test.

The machine can only make probabilistic and uncertain decisions.

Ultimately though, focusing on the algorithms may be a mistake altogether. One may argue that you do not need to find the perfect solution. There may be no need to be find a solution more precise than the data from sensors.

If you know your vehicle’s position only within 1mm, whether there is a solution you found that is 1mm better is beside the point. In most real-life cases, the best option is to brake.

Many consider that achieving Level 4 autonomous driving in NYC, where driver eye-contact and pedestrian body language are paramount, to be the holy grail of self-driving AI.

The “safest” autonomous technology may not necessarily mean the “best” autonomous technology. In some cases, we may want to work towards making the autonomous driver indistinguishable from the human driver, the ultimate Turing test.

Though these problems are fun to think about, they detract from trying to solve Year 1 problems in safety, legal liability, and insurance. While CalSTA and DMV may require reports on mileage and disengagement, these safety precautions are mostly surface level.

System or code malfunction can occur without any disengagement detection. No one will monitor the underlying risk except for the engineer operators themselves or the eventual human pilot.

Any product liability expert will share that these “defects” will not be known to the public until an aggregate number of accident happens.

The purpose of proper insurance is to ensure that these unknowns are held accountable for, and we argue for more public commenting on this subject.

Training of Machines

James May enjoying a past time. Top Gear, Season 1, 2002.

The largest disservice the autonomous community has committed against themselves is pitch policymakers the same way they have pitched their investors.

The term “autonomous” confuses the public and government officials. We all want to achieve the status of an autonomous vehicle, but all R&D vehicles are not “autonomous” from the start or by simply labeling the car as so.

Today, we simply have R&D, test, and prototype vehicles that need to go through substantial training. The development of deep learning and neural net frameworks must be put through hundreds of millions of miles of training.

Sensor arrangements and drive-by-wire systems must be extensively tested.

Furthermore, each manufacturer needs to rigorously test edge case and fix bugs that could be caused by changing weather conditions, new operating domains and road conditions, and unknown unknowns created by behaviors created by human drivers or human actions.

The difference of this new generation of R&D, test, and prototype vehicles versus those of previous generations is that the former requires machine learning in live road conditions, whereas the latter could be accomplished in closed course facilities.

However, the nomenclature or phases of R&D, test, and prototype deployment vehicles are barely defined by the manufacturers themselves, let alone regulators, as each manufacturer may assert confidence in a phase just to curry more investor interest and consumer demand.

In truth, every manufacturer will have a certain confidence score in an operational design domain. High confidence scores will be asserted, but with a non-statistically significant sampling.

We should understand that for the next 5 years, all these machines, even at the very late stages of commercial deployment, will be continuing to train and improve.

Manufacturers may want to just train their technologies passively with an active human pilot. Ultimately, a natural person in some capacity should always be responsible and liable during the training of the machine.

The automobile pilot may take form of a person behind the wheel, a person with a controlling laptop in the back-passenger seat, or a team of persons in a pursuit vehicle or distant command tower.

Nevertheless, there will be a human controlling these layers of systems, even if it is an idle engineer behind a screen of terminals monitoring system and individual unit functions, and that is where responsibility is shouldered in terms of driver liability.

If the public and government officials understand this, then we can propose a very pragmatic solution that is win-win-win for policymakers, manufacturers, and consumers.

As for insurance, the business case may shift between commercial auto and personal auto. Some companies want to take all liability.

Experienced veterans that see cyber risk unfolding today, make the assertion that technology change usually has implications on the entire insurance produce line and value chain.

We believe the faster the insurance and lending industries understand this, the faster we can scale testing.

Borrowing from Aviation

#34 Wilbur Wright, 1910.
Early Aviation Photograph Collection. Copyright 2009, President and Fellows of Harvard College.

The pragmatic solution stems from the aviation industry, since commercial airlines today are mostly self-flying, though no one in the aviation industry actively advertises this “autonomy” to the public.

Regulators have an easier time setting the rules of the sky than the rules of the road, because it is far easier to regulate two major manufacturers and spot a safety problem when a plane falls out of the sky than it is to regulate a growing list registered manufactures in California.

Two guys and a dog with no more than $500k of funding can claim to have an autonomous vehicle and no one will know of a serious safety failure during testing out in Bakersfield if the owners do not self-report the flaw, let alone self-detect the flaw.

Whereas in aviation, the matured industry has led to consolidation and the layers of redundant fail safe systems as well as safety procedures even before one is sold to a commercial operator, let alone when it is commercialized and an end-consumer becomes a passenger.

While it is easy for policymakers to hitch onto the innovation wagon and move along rules that appease Silicon Valley or Detroit, this becomes a nightmare for those who become financially responsible, the insuring parties.

The question of liability looms large, but can be easily solved with smart legislation that follows the aviation industry, the pilot-command concept, and tremendous aviation case law and legislation.

The financial responsibility of the manufacturer, the vehicle owner, and the pilot are clearly delineated from the customer.

Simply, a company such as Boeing bears responsibility of the product it manufacturers, a company such as United Airlines implements the Contract of Carriage, and the trained aviation pilots (no matter how “autonomous” the jet becomes) have proper training and override mechanisms.

While many more allegories and lessons can be drawn from aviation, the purpose here is to illustrate that the rate limiting step is not the lack of human IQ or deep learning brainpower, and it should not be regulation or lack of insurance, but taking the first game-theory step in cooperating and addressing the common safety standards and goals.

We need to ensure that policymakers, financiers, insurers, consumers, and the press, can work alongside manufacturers, to shorten the time horizon by a factor of 10.

The main issue for CalSTA and DMV here is to stretch their problem-solving approach and adopt ideas from FAA and NASA. The latter two agencies put substantial energies thinking about preventing catastrophic events.

The terminology “Critical Event Control” borrowed from British barristers and legislators who are promoting their own AV industry may be useful for our own purposes.

Whereas with aviation there are today requirements for airplanes horizontally, vertically, laterally, and distance between jets, there are none today for traditional vehicles.

The difference between aviation and the road is that when something goes wrong in the sky, redundant warnings and safety systems gives the pilot minutes before minutes to recover.

Even if there is a sudden catastrophic failure in the aircraft, the aerospace industry has put in effort to collaboratively put in procedures and standards. Things on the road happen in split seconds.

Addressing Critical Event Control is paramount for insurers and should be for regulators who have public safety in mind.

And here we conclude with the beginnings of aviation: the Wright Brothers. While they lived in an era with no regulation and no insurance being required in Kitty Hawk, North Carolina, and whether these two guys even had a dog to fit the startup founder archetype, it is certain that they did make their major achievement on American soil, and the industry found fertile ground for dreams to take off.

Thank you to the former professors, industry veterans, mentors, and more than a dozen lawyers across multiple specialities based here in the U.S. as well as the U.K. for providing feedback on this letter.
Albus Insurance is a commercial auto MGA developing a program specific to autonomous technologies. Reach me by clicking here: Christopher Lee.
Like what you read? Give Christopher Kelvin Lee a round of applause.

From a quick cheer to a standing ovation, clap to show how much you enjoyed this story.