Autonomous Vehicles (AVs) and the Double-Edged Sword of Safety Data

In one of the latest innovations coming from Tesla, their vehicles division recently launched a beta program called the “Tesla Safety Score”. The basis of this score is an algorithmic breakdown of several factors that are recorded while operating the vehicle. These various data points are all fed into a formula that results in a 0-100 type score, ranking how “safe” of a driver you have been since the program started. While this blog will focus primarily on Tesla’s implementation of this score and its implications, I want to clarify that I believe this is only a very early example of how widespread consumption of AV data will change not only how we drive but also insurance companies and other related industries determine per-driver rates.

What is the Tesla Safety Score, and how is it calculated?

Tesla Safety Score in the Tesla App

Tesla brands this score as an “assessment of your driving behavior based on five metrics called Safety Factors”, which are then all taken into consideration in order to provide a “safety rating” likelihood that your driving habits could result in a future collision. This assessment utilizes the past 30 days of driving data and aggregates your score. The ultimate intent of the score is to provide a deeper level of transparency and feedback to drivers. Let’s take a deeper look at the five metrics curating a driver’s score:

  1. Forward Collision Warnings
    1. Frequency of triggering a warning when about to collide with a foreign object ahead of the vehicle (per 1,000 miles)
  2. Hard Braking
    1. Backwards accelerations in excess of 0.3g. Hard braking while Autopilot is engaged does not factor into this variable
  3. Aggressive Turning
    1. Force of left/right acceleration exceeding 0.4g. Aggressive turning while Autopilot is engaged does not factor into this variable
  4. Unsafe Following
    1. Defined as the proportion of time where your vehicle’s headway is less than 1 second relative to the time that your headway to the vehicle ahead of you is less than 3 seconds
  5. Forced Autopilot Disengagements
    1. When the system disengages for the remainder of your trip due to receiving 3 audio and visual warnings

These factors are then calculated into the following formula. This formula determines a PCF, which is then converted to a 0-100 score based on statistical modeling using 6 billion miles of fleet data. The formula is predicted to change as more data is compiled globally:

With the resulting Safety Score, Tesla currently has two primary uses for the metric: vetting out Beta testers for the next iteration of Full Self Drive (FSD) and provide more driver specific discounts through their own internally provided vehicle insurance.

Releasing the next major iteration of FSD has been a highly anticipated feature that Tesla continually teases. Rollout plans were finally released along with the release of the Safety Score. Due to the massive amounts of Neural Network data FSD requires to continually learn and improve, a multi-month beta release has been planned. In order to gather the most reliable data from “ideal drivers”, the safety score has been introduced to identify the most ideal beta FSD testers. When a driver achieves a predetermined score, they will “unlock” the beta FSD upgrade that they can then download to their car. The exact requirements for unlocking the Beta are scheduled to lessen over time, with requiring a score of 100 for the first few weeks and then gradually lowering the requirements to open up the beta to a larger driver network.

Tesla UI prototype

In terms of insurance implications, Tesla has already taken steps to utilize this data to provide their insurance customers with more data-tailored discounts and benefits. While only currently available in California and Texas, Tesla has plans to roll out their insurance offering to many more states in the future. Specifically in Texas, they are able to roll out a program that will take Safety Score data to fine-tune insurance costs on a per-vehicle basis. This will also enable Tesla to proactively communicate any driving adjustments that could decrease the probability of collision. This sort of telematic based insurance is not the first to hit the market however, as insurance companies like Allstate currently offer data-based cost incentives when customers install approved telematic devices in their vehicles. The difference with Tesla’s implementation is that this data is already fully captured inherently within the vehicle’s storage. Any future AV/EV will already possess the technology needed to make telematic insurance rates a standard product offering.

Well that all sounds great… what could go wrong?!

With great leaps in technology, comes great risks in how that technology is utilized by end-users. Only weeks after releasing the Safety Score system, Consumer Reports came out with a statement claiming that the feature made users more of a liability on the roads. Tests performed at Consumer Reports resulted in scenarios where drivers would need to unsafely drive through intersections or not stop completely for pedestrians in order to maintain a low “hard braking” score. While these seem like easy tweaks, they bring up the revelation that there will always be real-world scenarios that go unaccounted for. However Tesla decides to model the “safest driving” scenario, the understanding needs to be socialized that a level of real-world common sense must always take center stage.

Ultimately I do believe that this sort of data usage from AVs is the future for roadways. I think there is too much benefit to consumers and retailers alike to have quantifiable data to show driving trends, which could then lead to optimal insurance rates and other similar cost reductions. I also believe that this sort of technology usage on the roadways could be analyzed at a more holistic level to understand driver trends and implement more effective traffic patterns etc.

Only time will tell how effectively we use the almost unlimited sources of data now at our fingertips.

Sources:

https://www.tesla.com/support/safety-score

https://www.notateslaapp.com/news/591/a-look-at-tesla-s-safety-score

https://www.repairerdrivennews.com/2021/10/27/draft-tesla-insurance-unleashes-unique-safety-score-program-in-texas/

https://www.teslarati.com/tesla-safety-score-unsafe-consumer-reports/

https://www.tesla.com/modely/design#overview

17 comments

  1. This was a fascinating deep dive, thanks for posting, I had no idea they were doing this. I wonder if we’ll look back at this in a few years when Tesla is a dominant insurance player, as a perfect example of how platforms and data enable new entrants to seemingly unrelated markets. A quick search shows me other 3rd party bodies like FICO are working on similar types of solutions. Since all cars have a lot of this info now, whoever can connect to all the different systems and create a product that can be used with a multitude of car companies and insurance carriers could clean up.

    1. Spot on. I think Tesla has the capabilities to really create the ideal data-gathering safety tool, but I don’t think there is any reason that a 3rd party or other insurance/car company couldn’t take this idea and market it well beyond just incentivizing a Full Self Drive beta invitation.

      My personal opinion with Tesla is that they spread themselves far too thin with overarching objectives, which really dilutes their ability to push out polished products. Their cars are incredible machines, yet they can’t get a hold of the Quality Control or customer service/support, which could very well come back to to haunt them long-term. I’m afraid this could fall into the same category of being a really good idea, but not getting enough focus in-house at Tesla.

  2. Could we see Tesla equip their own insurance for their cars soon? I don’t like the idea of a Safety Score as it could prevent drivers from that promotion. What if mom was driving and decreased the score you otherwise would’ve had. How can Tesla account for that, choose an account for who’s driving before stepping on the pedal? That’s how Netflix is supposed to work but we’ve all used other profiles on an account. Good idea, but it sounds like they have a lot to think about.

    1. Yeah they already offer their own insurance in California and Texas, so I do think that nation/global rollout is the ultimate target for in-house Tesla insurance. And based on the blogs / subreddits I follow, a lot of drivers have brought up the same exact concern around knowing the driver associated with those 5 factors’ data gathered. Teslas do currently have “per-driver” settings, so when I get in I change it from my girlfriend’s settings to my own personal profile. This is a manual process however, so it could still be gathering inaccurate data for the wrong driver. They do have in-cabin cameras that will supposedly be able to detect specific users while in the drivers seat, so we’ll see how long it takes to get that into production.

  3. Great post and interesting that you can access the formula for safe driving. Wonder if AllState is as transparent? Or if gaining access to the formula is as transparent as it seems? With Hertz announcing the acquisition of Tesla vehicles and a potential partnership with Uber, do you think Uber may implement a safe driving score using the Tesla program? I, as a customer, would like to know how safe my Uber driver is based on an algorithm like the Tesla one.

    1. Allstate apparently provides “personalized driving feedback” from the on-board device you need to connect to your vehicle, but I don’t think it is nearly as in-depth as the Tesla metrics at this point (https://www.allstate.com/tr/car-insurance/telematics-device.aspx).

      And regarding the acquisition, I think this is definitely one of the major use-cases for the safety-score. Assuming they are able to refine the metrics well enough that it does accurately define the right drivers as “safe”, I think that will provide them Uber with a massive advantage over any other ride-share offering. Being able to see the specifics of your driver’s score (if marketed and implemented appropriately) with transparent metrics and data could be an absolute game-changer for Uber and Hertz.

  4. Part of what has always fascinated me about tech is unintended consequences. For instance, many “no text” laws actually increased accidents, because people were just texting below the window lines so they wouldn’t be seen doing it, which created much less safe conditions

  5. This makes me thing of the conversation we hard about AI the other week, in that assumptions and biases are built into the program itself. I am curious to see how this data informs currently existing stereotypes regarding the gender and ethnicity of drivers. This Forbes article highlights that en often pay more for insurance in the earlier driving years while women tend to pay more later https://www.forbes.com/advisor/car-insurance/rates-age-and-gender/. This article highlights the discrepancies in car insurance rates in neighborhoods where people of color live https://www.thebalance.com/why-do-minorities-pay-more-for-car-insurance-4163553#:~:text=Yes.,not%20just%20predominantly%20Black%20neighborhoods.

    I would be curious to see how this research lines up with what Tesla is bringing out (seemingly transparently)!

  6. Great stuff! I can see it now, “Tesla best in safety and best for the environment!”

    I am curious what happens if more than one person drives the vehicle during the 30 day data gathering period. How it would effect the safety score? Which leads me to another question, if and when Tesla officially releases their Full Service Driving will there be any sort of additional driving test for owners of autonomous vehicles. Furthermore, will there be different insurance rates for autonomous vehicles versus human driven vehicles?

  7. Awesome topic, Bryan! I appreciate how in-depth you went. This is yet another example of the vast amount of data collection taking place every moment of our lives.

    I’m not much of a car person, so I wonder how Tesla’s safety tests compare to other companies like On-Star? From the short research I did, they all sound the same.

    Side note: I can only imagine what my test results would say haha…”dances too much while driving causing high speeding”

  8. Great job Bryan! Love your point about the insurance practices. I have to say in the past I got burnt out pretty hard the Progressive app that was giving me $15 a month to allow them to track my data, turns out I am not what they consider a good driving for their standards. I like to go fast and brake hard! All jokes aside it was due to the app being unable to actually do a reliable job tracking my actual rides. I complained about the app and then I end up switching companies. I believe Tesla has the right technologies and expertise to make this a reality. Thanks again for your great blog!

  9. Thanks for sharing this information. I find a lot of parity with Progressive’s Snapshot Rewards/Safe Driver Discount. Instead of being an application or actual device, you plug it into your car, and it now comes built into your Tesla. This makes a bit more sense since Tesla has the sensors to track more complex interactions of the driver. I’m interested to see how insurance companies will begin to use this information to update their policies.

  10. As Parker touched on, I wonder if Tesla will find a new revenue stream through selling the software (used to calculate the Safe Driver Score) to other driving companies. If Uber acquired such tech in the future, the ride-sharing platform would most-likely secure the top spot in the ride-share industry due to the renewed sense of safety. Customers would feel much more protected seeing a tangible driving score (rather than the 5.0 rating that most Uber drivers acquire). The Safe Driver score would also be beneficial for Budget Truck rentals too. I always see the sticker on the back of those trucks that say “Call XXX-XXX-XXX to tell me how I am driving!”, but with this tech, their driving skills would already be calculated for Budget Review.

  11. Finally, some electric car discussion! I had no idea this feature was being driven out, pun intended, consequences un-intended. Various 3rd party tools have been created in the past to measure driver safety such as the Allstate safe driver program (https://www.allstate.com/drive-wise.aspx) and breathalyzers like IntoxaLock (https://www.intoxalock.com/). I tend to be adverse to the insurance company marketing which claims the business is pushing safe drivers for the sheer betterment of the objective vs. their true intentions of saving more money via less insurance claims. I’m glad the consumer reporters stepped in to audit this feature and definitely interested to learn more on the ramifications of this feature & the selling of its data.

  12. Nice post Bryan! I like your point at the end about how this data could be used to impact transit decisions and development. I think this data could help transit planners better identify issues for example I wonder if the car can report back data based on location. Maybe there is a part of your commute where you are prone to hard breaking. Perhaps data from other cars might show a similar trend establishing a pattern. This might help a transit planner identify an area for improvement in the roadway at this location.

  13. Bryan, I love how you joked in your presentation about not being an insurance expert, but keep bringing it up. I think you’re becoming the class insurance expert! And if so, consider me your sidekick! I want to comment on your mention of “telematic-based insurance” and how Tesla is building it into their cars. In my insurance law class, we discussed how insurance companies’ usage of risk classification can lead to discrimination (if X people are Y% more likely to have car stolen, then they get charged Z% more than XX people). The goal of insurance companies, when using these tele-metrics, is to skim off the low-risk customers, i.e. discriminate against them. Persuading soon-to-be, or soon-not-to-be, policyholders to wear something to measure daily habits (like these car ones that measure your speed, how hard you brake, etc.) and then streaming the data real-time to insurance companies who will raise or lower rates based on how safely you drive can easily be discriminatory — against, perhaps, teenage boys, etc. In driverless cars? Tesla will be ensuring that their technology matches the standards required of these insurance companies to get low rates for who?? Their drivers or their cars/software behind autonomous driving? The problem is going to be what happens in the event of an accident when the car, not the driver, was driving. It’s unclear what risk the insurance company will be insuring based on this Tesla example, but I’m certain that insurance companies will be making these decisions very carefully (and not for the benefit of Tesla, or their drivers).

  14. I find it scary that all the data is intrinsically collected by the car itself. I believe that in the near future there will be no way to escape legality ( this should be “good” )
    You may receive a fine that your own car reported to the police.
    I hope that in the long-term all this data control reports more benefits and that as users we are willing to give up on that privacy for shared wellbeing. After all, there are millions of people that die in a traffic accident and I am willing to sacrifice privacy to avoid human loss.

Leave a Reply

Please log in using one of these methods to post your comment:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: