On the emerging risks of automation: the case for Autonomous Vehicles
The future of mobility is autonomous. The change is already well underway. Cars have started assisting our driving and will eventually be driving themselves. Over the next fifteen years, we will be heading to the permanent establishment of autonomous vehicles (AVs) on our roads.
Article information and share options
It will start small. The deployment of AVs will initially be limited to selected areas before expanding in and outside of cities (see Figure 1). The impact on society will be significant, as it will be on the automotive industry and, in turn, on the re/insurance market1.
In this article we provide insights on some of the risks that may emerge, highlighting the need for re/insurers to understand AV technology, as re/insurance solutions will gradually move from being driver-centric to vehicle-centric. Data analytics capabilities will be key in this process and with it the need to create collaborative ecosystems with tech providers and original equipment manufacturers (OEMs) to exchange knowledge and develop tailored insurance solutions.
As the deployment of AV is still at an early stage, AV specific re/insurance products are still being conceived and developed. This publication attempts to fill some gaps on the subject, by providing unique perspectives. Making predictions about the riskiness of AVs is a speculative and complex exercise. In fact, the purpose of this publication is not to forecast, rather to provide a conceptual framework to understand the effects that AVs may have on the re/insurance industry.
Figure 1: AV deployment over the next fifteen years
Understanding AV technology
Most new vehicles now include some form of Advanced Driver-Assistance System (ADAS). The technology assists and supports drivers in emergency interventions, such as emergency braking, and/or comfort enhancement, for example cruise control. The operational boundaries between safety and comfort features are often blurred as, typically, sensors are shared across both functionalities.2
The presence of ADAS technology does not mean these cars are equipped with self-driving capabilities. An autonomous vehicle (AV) is conceptually and technically very different to one equipped with ADAS. Conceptually, AVs (unlike ADAS vehicles) are conceived as a whole unit: a software-driven full-integrated body. Technically, the use of Artificial Intelligence (AI) allows the AV to learn from situations, make decisions that might closely resemble human choices, and even enhance their typical expected performance3.
Vehicle autonomy is typically ranked from level 0 (no automation) to level 5 (fully autonomous everywhere and in any conditions4). Throughout this text we refer to traditional vehicles (with or without ADAS) as those in the 0-2/3 levels and autonomous vehicles as those in the 3/4-5 levels.
Table 1: Levels of vehicle automation vs driver expected engagement
While the shift from traditional to autonomous vehicles cannot be easily assigned to a specific level, the transition from level 2 to 3 is generally recognized as an inflection point, at which ADAS technology changes over to increasingly integrated and automated systems. In a level 2 vehicle, the driver must still pay attention at all times. In a level 3 vehicle onwards, the driver can afford to relinquish control, even if only under specific circumstances. The scope of these circumstances increases with level 4 and control is totally relinquished with level 5.
Traditional car manufacturers are following a sequential strategy of developing vehicles, proceeding from Level 1 to 2 and onwards. New entrants, dominated by tech giants and tech start-ups, are seeking to leapfrog intermediate steps to reach full autonomy earlier. These divergent approaches are driven by different business/commercial priorities (target segments and use cases), strategic considerations (long-term visionary approach), and technical choices (modelling human-machine interaction, which is an inherent and highly complex property of intermediate levels of autonomy).
Software in autonomous vehicles
Software in AVs needs to process large volumes of data to understand internal and external environments and, subsequently, take optimal decisions. Sophisticated hardware and software components need to work smoothly with each other. The complex interplay of different algorithms makes it hard to dissect software into functionalities and associate a purpose to each part. Nevertheless, the literature5 indicates four main units of every AV's software stack: perception, prediction, planning, and control.
- The perception module processes raw sensor data to generate an actionable representation of the environment for decision making. This involves tasks such as object detection, 3D reconstruction of surrounding environments (mapping – HD maps), localisation and state estimation. The state-of-the-art perception modules make extensive use of machine learning methods for computer vision tasks.
- The prediction module predicts the future behaviour of other traffic agents. For example, it creates a likelihood of possible future trajectories of another vehicle at a crossing in order to avoid collision.
- The planning module determines navigation routes based on the representation of the environment obtained from perception and prediction. One can usually distinguish between higher-level decisions (e.g. whether or not to change lanes) versus lower-level ones, such as finding a suitable trajectory for the lane change.
- Finally, the control module sends the appropriate signals to the hardware components of the vehicle, the throttle, braking or steering torque to follow the planned trajectory.
Hardware in autonomous vehicles
AVs require a variety of sensors to (a) map and understand their surroundings and to (b) assure redundancy, so that sensors' field of view overlap in the case of failure or malfunctioning. Sensors can be classified into two types, exteroceptive (cameras, LIDARs, RADARs) and proprioceptive (GNSS, IMU). The first are used to gather information relating to the external environment, while the latter are used for information about the relative positioning of the car6.
Figure 2: AV's key hardware components
Cameras are used to capture images of the surroundings and create a 3D view of the environment. However, they are not ideal for depth perception and do not work optimally in conditions of diminished visibility, rendered either by weather or daylight.
RADARs detect speeds and distances accurately and, most importantly, work in low visibility conditions. However, they are not able to precisely identify the shape of an object, so they might have difficulty distinguishing the shape of, for example, a human from an animal.
LIDARs (light detection and ranging) detect objects and measure distances by illuminating the target with pulsating lasers and analysing reflected light. They are considered, by most AV providers, one of the key components of an AV, able to generate high precision 3D maps of the vehicle surroundings. They also have limitations in low-visibility weather conditions.
GNSS (Global Navigation Satellite System) assists the vehicle with accurate localisation and is typically integrated with an inertial measurement unit (IMU), used by the vehicle to determine its orientation and angular rate.
Connectivity enables the vehicle to communicate bidirectionally with other systems (vehicles or infrastructure) outside of the car. Connectivity is not essential for AVs to operate, but it could enhance performance and safety by providing additional contextual data points for decision making. For example, connected sensors may indicate whether the vehicle ought to move, thus avoiding stalling situations where the vehicle is uncertain what to do or perceives that the risk is too high to continue.
We can distinguish three types of vehicles connectivity:
- Vehicle to Everything – V2X allows the vehicle to stay connected with its surroundings, which could include other vehicles (V2V), road or surrounding infrastructure (V2I), or local networks (V2N).
- Vehicle Telematics allows information to flow to and from the vehicle. Its role ranges from enabling map and software updates to car signals transmission.
- Infotainment (in-car entertainment) is already present in many traditional vehicles and its adoption, not related to the functionality of the AV, is expected to increase with the rise of AVs, as the user, no longer required to drive, can engage in other recreational tasks.
Industry dynamics and emerging risks
New industry dynamics and solution structures
AVs will affect every stage of the automotive value chain, changing how vehicles are designed, manufactured, sold, owned and serviced. These changes could have important implications for insurers.
Figure 3: Value creation via the 'smiling curve' concept
The smiling curve, a concept first developed by Stan Shih in 19927, provides a framework to characterize how AVs might change the automotive value chain. According to this framework, the stages at the extreme ends of the supply chain (those closest to and furthest away from the customer) will add the most value. For traditional vehicles, these stages are R&D and sales service. The biggest added value for AVs will be the development of their software and hardware components. Companies specializing in these production stages will likely hold the strongest bargaining power in the market8.
Due to the deep technical expertise required for production of AVs, many well-funded tech start-ups and larger tech corporations have been entering the automotive market. Traditional car manufacturers have been quick to develop strategic partnerships with these new players (see also: How demand and supply are moving closer to equilibrium). These newer entrants will also be important for insurers as they will control substantial volumes of data, access to which could prove important for the optimal assessment of risks.
Players active in the AV industry structure their solutions in different ways. We have highlighted two approaches that we think might be relevant from an insurance perspective:
- Fully integrated solutions, where a single company develops the AV technology jointly with the vehicle, as to enable a full integration between the two. This allows for improved calibration of software, hardware (sensors), and the expected behaviour of the rest of the vehicle. Not all components will be developed in-house (some hardware components are often outsourced), but optimal interplay between the various elements of the vehicle is inherently attained.
- AV technology providers will only develop and supply the AV technology (software) which they will subsequently sell to car manufacturers and/or fleet operators. The algorithms will not be calibrated to any specific hardware and it will become the fleet owners' responsibility to ensure good operation of the AV technology (even though we expect close collaboration between the software provider and the car manufacturer).
From an insurance perspective, fully integrated solutions will likely carry a lower risk. The consolidation of all stages of production under one company will produce calibrated software and hardware, ensuring improved predictability and optimality of output for a given set of inputs.
Emerging AV risks
The adoption of AVs will lead to the emergence of risk types not present (or just partially present) in traditional vehicles. In this section we provide an overview of these risks (Figure 4) and their possible evolution with levels of autonomy (see Table 2).
Figure 4: Emerging risks in AVs
We classify AV risks in four types:
1. Behavioural risk
- Driver risk: In a traditional vehicle this risk is mainly driven by human error, usually attributable to (a) intentional misconduct and reckless behaviour or (b) distraction/ tiredness of the driver.
- AI risk: Driving tasks are embodied by the AI. This risk is hard to assess due to: (a) difficulties in gaining access to the source code and (b) deciphering and making sense of such code. Predicting the behavior of AI in different driving scenarios is not straightforward. On top of this, comparing the risks of different algorithms (e.g., assessing which one leads to the safest set of solutions) inter- and intra-provider can be an arduous task. A proxy for this clear box approach (accessing and studying the software components and their performance) might be a black box approach, studying the output of the AI in terms of a vehicle's behaviour, ignoring the underlying roles and tasks of software and hardware components9.
2. Internal risk
Internal risk refers to the possible failure or malfunctioning of the vehicle's hardware, both standard parts (e.g. brakes, steering wheels) and those that enable automation (e.g. radars, cameras). Redundant systems play a key role in ensuring that the vehicle can guarantee performance in circumstances of malfunctioning or failure. For example, the field view of a radar could overlap with that of a camera, thus having two sensors covering, a similar area. Internal risk can also be related to software failure, for example a mismatch between the algorithm and the software version.
3. Connectivity risk
If networked driving becomes key function of the AV, network failures will give rise to connectivity risks (see section on vehicle connectivity). For example, a vehicle that relies on the cloud to complete software updates could risk a drop in performance associated with a drop in connectivity.
4. Cyber risk
Malicious cyber-attacks could affect all three of the above risks. Most malicious attacks fall into the following categories:
- Intention to steal the vehicle (already present for traditional vehicles)
- Take control of the vehicle for ransom
- Theft of (personal) data
- Disable functions so that the vehicle crashes and/or no longer moves
As the level of autonomy and connectivity progress, the attack surface increases and consequently so does the risk of malicious attacks. While this is important to anticipate, we are yet to see large scale cyber-attacks on connected vehicles. Automotive companies and enterprises are increasingly investing in cybersecurity as the automotive space becomes increasingly software-reliant and connected10.
Level 1 to 4 vehicles will experience, to a varying extent, an element of HMI risk (Human machine interaction)11, which falls into two broad causes/origins:
- The human driver fails to react quickly and appropriately to the warning signs of a vehicle's ADAS. This is often due to poor understanding of the technology by the driver and/or (lack of) learning to properly use it.
- The human driver engages in overconfident behaviour because the vehicle is equipped with ADAS.
Table 2: Possible sources of risk at different levels of automation
AV risk might also vary as the objective functions that steer the decision making of the vehicle may vary. For instance, delivery AVs may decide to assign greater weight to a "progress towards the goal" (i.e. speed) rather than comfort, depending on the fragility of the cargo, while a school bus AV may prioritize comfort and safety at the expense of speed.
We expect early AV providers will be unlikely to tweak optimisation objectives based on use cases. It is cheaper, more convenient and simpler to maintain a single testing system over doing so for multiple ones.
The implications for insurance
The challenges of assessing AV risks
While the increased adoption of ADAS technologies has been slowly changing road traffic dynamics and related risks, it is harder to assess the risks associated with AVs for four main reasons:
- Their intrinsic nature: The behaviour of AVs is an emergent property of various software and hardware units that cannot easily be reduced to any single component12.
- Opaque decision-making: It is hard to single out which part of the AV source code does what and, therefore, define deterministic and unique behavioural specifications. Even for coders, the learning and reasoning processes of black box neural networks are not obvious.
- Non-linearity of decision-making: The perception, prediction, planning and control functionalities of AVs do not work in a linear and sequential way, but rather in an asynchronous and highly intersected manner.
- Exhaustive definition of Operational Design Domain: AV riskiness might be dependent on the definition of their ODD. However, even within a specific ODD, it may not be possible to consider the full extent of scenarios' characteristics that an AV encounters.
Further, non-tech-related, difficulties will possibly stem from:
1. The assessment of intentionality of the actions executed by AVs will be difficult to ascertain; and, if so, even harder to attribute responsibility. This has important implications as current liability policies exclude "intentional acts" from their coverage.
2. An initial phase in the adoption of AVs where liability is split between the different parties involved, rather than a clear and unique assignment of fault. In this regard, it might be worth distinguishing between two types of scenarios:
- Traditional vehicle vs AV: We speculate that initially AVs will be found, on average, more often "liable" than traditional vehicles. With time, as the performance of AVs will improve and so will trust in their technology, the situation might be reversed. In the long term, AVs may thus become more attractive to insure compared to traditional vehicles.
- AV vs AV: In this scenario, it could be difficult to assign responsibility, which may lead to a 50/50 split of liability. This could disincentivise providers to improve their technology, thus leading to a possible levelling down in performance.
3. Deep pocket issues, namely the fact that AV/tech companies, as they are often well capitalized, might become the target of lawyers for expensive suits in the case of a cluster of accidents and unclear liability assignment.
4. The legal determination of a vehicle is internationally embodied in Article 8 of the Vienna Convention (1968), stating that "every moving vehicle or combination of vehicles shall have a driver;" and that "every driver shall at all times be able to control his vehicle.13" National and subnational legislatures have implemented changes allowing AVs in at least a trial capacity, but legal traffic frameworks are still structured around the notion of human control of a vehicle and will require further amendment.
Predicting the impact of AVs on insurance-related aspects is not at all straightforward. Here, we provide some insights on the possible effects for insurance rating factors, frequency, severity and claims management.
Existing motor underwriting models will probably not experience significant changes. However, the explanatory variables used to assess risk may be substantially different as a result of:
- the presence of additional and different types of data
- the gradual or abrupt shift in task execution from the driver to the vehicle
Insurers will gradually need to move from assessing the risk profile of the driver to the risk profile of the AI.
The breadth of AV use cases could impact insurance rating structures in a more nuanced manner and to a greater extent than in the case of traditional vehicles. Whether the vehicle is employed for the transportation of goods or human beings already has an impact on insurance pricing for traditional vehicles. However, this difference could be even more marked for AVs for two reasons:
- the objective function of the algorithms might be tweaked based on what is being transported; this distinction in the algorithms based on use cases is likely to be a long-term strategy for providers (see section above on AI risk).
- the pricing could depend on the amount of risk the customer is willing to retain, which will probably be higher for goods than people (with manual delivery, a driver is always present).
Nevertheless, large motor losses result from bodily injuries to vulnerable road users (e.g., pedestrians, cyclists): a heavy AV carrying goods might cause very severe bodily injuries to such road users.
Furthermore, a better quality and higher frequency of the data will allow for enhanced product tailoring. This might enable a more precise assessment both in terms of real-time risk (when and where a vehicle travels) and risk profiles of different vehicles' makes and models.
Changes in accident frequency
Current data on the frequency of AV accidents is limited, as most AVs are deployed and tested in controlled environments. However, two time-dependent use cases can be considered:
The initial stage of AV adoption might have a negative impact on accident frequency due to the interaction between traditional vehicles and AVs. We expect that AVs will maximise adherence to traffic rules and safety, thus making them, from a human standpoint, overly cautious. This might lead to an increase in accident frequency due to unexpected behaviours such as sudden braking. Moreover, AVs might underperform ADAS equipped vehicles in specific functions. Indeed, the latter are programmed to engage in precise tasks, as opposed to AVs, which are conceived to emulate (and arguably improve) driving behaviour as a whole. On the other hand, regulators will demand that AVs are safer than average drivers in traditional vehicles.
In closing, we expect that the overall balance on frequency will be positive (i.e., reduction in claims frequency) as the impact of human error reduction might outweigh their unexpected behavior.
AVs will become predominant on the roads and the negative impact of unpredictability will diminish. Connectivity will also have a positive impact on the frequency of accidents, as AVs will learn and have the capability to interact with other AVs and surrounding infrastructure.
We expect, overall, a more pronounced positive effect through the reduction of accidents in the longer term.
In a similar manner to accident frequency, the severity of material damage will depend on the time frame of AV deployment. Initial cost loads may be high, reflecting expensive component costs and a scarcity of trained mechanics. The costs of both will come down in time as the technology becomes mass produced.
As for the severity of bodily injury, this is likely to remain relatively stable with the emergence and adoption of AVs.
Although signals from AVs (e.g., accelerations, sensor outputs) may be used to assess damage after a collision, rapid access to these signals is mainly due to better connectivity in the vehicle rather than any autonomous capability of the vehicle.
Tech-savvy insurers might already have the data and capabilities to pay claims in (nearly) real time using existing visual intelligence (e.g. photos via mobile phones) and computer vision- based machine learning. Moving forward, as vehicles become more connected, the usage of car data in connection with visual intelligence and machine learning will gain momentum and significantly improve the accident reconstruction process, the assessment of damage, and the speed of claims payments, which is key to profitability. In particular for AVs, more than for traditional vehicles, accident reconstruction will be pivotal in assigning fault responsibility.
Technology will lead the way in the delivery of AVs. Nevertheless, its adoption rate will mostly depend on the attitude of AV providers, consumers and regulators rather than technology. It will largely boil down to questions of trust; and with transparency, vigilance and continuous improvement, this trust can be created.
The ability of re/insurers to support the AV industry will strongly depend on their capability to access the right information, understand the underlying algorithms, and predict AV behaviour. AVs are more complex than ADAS-fitted vehicles and developing strategic partnerships with the right players will become essential in achieving these goals.
After an initial transition phase, AVs will ultimately mean safer roads. Like any automated technology, they will not be risk free; they will still need insurance. However, on average they are expected to be safer than human drivers. And from an insurance and a societal perspective, that is a development we can only welcome.
The authors would like to thank Jonathan Anchen, Head Swiss Re Institute Research & Data Support, Marc Freuler, Senior Underwriter Facultative Casualty, Andrea Biancheri, Senior Pricing Actuary, Alessandro Zanardi, PhD Student, ETH Zurich, for their contributions to this digital publication.
- 1 N. Lang et al., "Can Self-Driving Cars Stop The Urban Mobility Meltdown?", BCG Global, 2020, https://www.bcg.com/publications/2020/how-autonomous-vehicles-can-benefit-urban-mobility; K. Heineke et al., 2017, https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/self-driving-car-technology-when-will-the-robots-hit-the-road; M. Winkler et al., "Accelerating Automotive's AI Transformation: How Driving AI Enterprise-Wide Can Turbo-Charge Organizational Value", Capgemini.Com, 2019, https://www.capgemini.com/wp-content/uploads/2019/03/30-min-%E2%80%93-Report-3.pdf.
- 2 L. Di Lillo, "The Significance Of Progress", Swissre.Com, 2021, https://www.swissre.com/dam/jcr:3b7ff782-ed6f-48ff-80f1-0a1193257697/veoneer-swiss-re-the-significance-of-progress.pdf
- 3 M. Mageira, "Autonomous Driving - An Interview With Luigi Di Lillo", Movingdots.Com, 2021, https://www.movingdots.com/news/are-self-driving-cars-safe
- 4 "SAE Levels Of Driving Automation™ Refined For Clarity And International Audience - SAE Levels Of Driving Automation™ Refined For Clarity And International Audience", Sae.Org, 2021, https://www.sae.org/blog/sae-j3016-update
- 5 B. Paden et al., "A Survey Of Motion Planning And Control Techniques For Self-Driving Urban Vehicles", IEEE Transactions On Intelligent Vehicles 1, no. 1 (2016), 33-55; A. Takacs et al., "Highly Automated Vehicles And Self-Driving Cars [Industry Tutorial]", IEEE Robotics & Automation Magazine 25, no. 4 (2018), 106-112; C. Badue et al., "Self-Driving Cars: A Survey", Expert Systems With Applications 165 (2021), 113816, doi:10.1016/j.eswa.2020.113816; A. Yoganandhan et al., "Fundamentals And Development Of Self-Driving Cars", Materials Today: Proceedings 33 (2020), 3303-3310
- 6 Orise.Orau.Gov,2021,https://orise.orau.gov/resources/stem/documents/scholarships/autonomous-driving-and-its-sensor-technology.pdf
- 7 "Smiling Curve - Wikipedia", Wikipedia.Org, 2021, https://en.wikipedia.org/wiki/Smiling_curve
- 8 R. Dingler, "Self-Driving Car Fleets: Transportation As A Service", Medium, 2019, https://medium.com/adventures-in-consumer-technology/the-new-business-model-of-self-driving-car-fleets-a14d94d61148
- 9 J. De Freitas et al., "From Driverless Dilemmas To More Practical Commonsense Tests For Automated Vehicles", Proceedings Of The National Academy Of Sciences 118, no. 11 (2021).
10 "CB Insights", App.Cbinsights.Com,2021, https://app.cbinsights.com
11 D. Plejic and S. Marcacci, "The Human Vs Machine Challenge With Modern Vehicles", Movingdots.Com, 2021, https://www.movingdots.com/news/the-human-vs-machine-challenge-with-modern-vehicles
12 J. De Freitas et al., "From Driverless Dilemmas To More Practical Commonsense Tests For Automated Vehicles", Proceedings Of The National Academy Of Sciences 118, no. 11 (2021).
13 "CONVENTION ON ROAD TRAFFIC", Unece.Org, 1968, https://unece.org/fileadmin/DAM/trans/conventn/Conv_road_traffic_EN.p