A Hard-Braking Count Is a Map of Where Somebody Lives
Usage-based insurance is sold as the fair alternative to demographic rating: we stopped guessing from your zip code and started measuring how you actually drive. The trouble is that most behaviour features are produced as much by the road as by the driver, and the road is the zip code by another name.
Why "We Only Measure Behaviour" Is Not a Defence
A telematics score is a model over sensor features, and each feature has a physical origin. Hard-braking counts rise with intersection density, pedestrian traffic, and stop-and-go congestion. Night driving is concentrated in shift work — healthcare, warehousing, delivery, hospitality. Road-type mix separates highway commuters from people driving arterial streets. Phone-handling inference depends on whether the car has integrated controls, which is a function of vehicle age, which is a function of income.
None of those features names a protected class. Together they can reconstruct a good part of one, which is the definition of proxy discrimination — and it is why a model can be built entirely from conduct data and still fail an outcome test.
Feature-Level Risk, Ranked
Hard-Braking and Hard-Acceleration per Mile
CRITICAL RISKDetermined largely by road environment. Dense urban and lower-income arterial routes generate more events for identical driving skill.
Night and Early-Morning Exposure
HIGH RISKA direct read on shift work. Penalises the occupations with the least schedule control, and correlates with income and often with race.
Phone-Handling Inference
HIGH RISKHandset-motion heuristics cannot always tell a driver from a passenger, or an integrated dashboard control from a handheld phone. Older vehicles score worse for the same conduct.
Trip Length and Frequency Distribution
MEDIUM RISKMany short trips reads as high exposure, but describes caregiving, multi-stop errand patterns, and gig work rather than risk appetite.
Total Mileage and Speed Over Posted Limit
LOWER RISKThe most defensible pair. Both are closer to driver choice and have a direct, explainable relationship to loss exposure.
Three Legal Tracks, Not One
Insurance rating law prohibits unfairly discriminatory rates and generally requires rates to be supported and not arbitrary. That is a filing-and-justification question, and it is where a proxy feature with no causal story becomes hard to defend.
Civil rights and consumer protection law reaches the outcome. A pricing pattern that lands disproportionately on a protected group invites the question of whether the same predictive power was available with a less discriminatory alternative — and with telematics there very often is one, because dropping a road-environment feature usually costs less accuracy than the modelling team expects.
Privacy law runs alongside both. Precise geolocation is sensitive personal information under several state regimes, and telematics collects it continuously, from a device that may also be logging passengers.
What a Defensible Telematics Programme Documents
Most of this is testing you should want anyway — the compliance value is that it exists in writing before somebody asks.
Audit your AI product's compliance exposure
RatedWithAI helps tech teams understand their compliance posture across accessibility, privacy, and AI regulation requirements. Start with a free scan.
Scan Your Product for Free →Frequently Asked Questions
The programme is opt-in and only ever discounts. Does that remove the risk?
It reduces it but does not remove it. A discount-only programme still allocates a benefit unevenly, and once enrolment is common the drivers who cannot earn the discount are effectively paying a surcharge. Regulators have shown interest in who is able to qualify, not only in who is penalised.
Our telematics score comes from a vendor and we do not see the features. Is that safer?
It is worse. You are still the party setting the rate, and 'the vendor's model produced it' is not an answer to a rate justification or a market conduct enquiry. Buy audit rights, feature documentation, and bias-testing results before you buy the score.
We do not collect race or ethnicity. How can we test for disparate outcomes?
Testing is done with an accepted inference method — a geography-and-surname-based estimate is the common approach in insurance and lending — applied to aggregate outcomes rather than to individual pricing. Not collecting the attribute is a reason to infer carefully for testing, not a reason to skip testing.
Is a driving score subject to adverse action notice requirements?
Where the score is supplied by a third party and used to decide eligibility or price, treat the consumer reporting rules as likely to apply and give a specific reason plus the source. Even where they do not clearly apply, a vague reason is the thing that turns a pricing complaint into a regulatory file.
What is the single highest-value fix?
Normalise the environment-driven features. Counting hard-braking events per mile of comparable road type, rather than per mile overall, removes a large part of the geographic signal while keeping most of the predictive content — and it is the change that is easiest to explain in a filing.