Tesla Self-Certified Its Cybercab as Safe. NHTSA Opened a Probe Within a Day.
Tesla began offering paid rides in its Cybercab, a two-seater with no steering wheel, no pedals, and no mirrors, on the streets of Austin this week. The National Highway Traffic Safety Administration opened a formal investigation into the deployment the next morning. What's notable isn't that a regulator is looking at a new autonomous vehicle, that happens routinely. It's specifically what NHTSA is investigating: not just how the Cybercabs behave in traffic, but the process and data Tesla used to self-certify the vehicles as meeting federal safety standards in the first place.
What NHTSA is asking
Federal vehicle safety rules generally require manual controls like a steering wheel and brake pedal. NHTSA's audit query, AQ26002, covers an estimated 1,000 Cybercab vehicles and targets Tesla's self-certification: the internal process and supporting data Tesla relied on to declare the vehicles compliant, absent a steering wheel or pedals, without going through the same path Amazon's Zoox took. Zoox operates a purpose-built robotaxi with no manual controls too, but it obtained an explicit federal exemption to do so before deployment. Tesla deployed on the basis of self-certification instead, without seeking that exemption route. NHTSA has separately proposed loosening the manual-control requirement for vehicles designed to be autonomously driven, which makes the timing of this specific probe pointed rather than routine: the rule Tesla may be betting will eventually change is the same rule regulators are actively investigating whether Tesla already complied with as written today.
Tesla's stock dropped on the news, and coverage has focused on the Cybercab's underwhelming launch specs alongside the investigation, which are separate but compounding problems for the same product.
Self-certification is a bet, not a guarantee
Self-certification exists across a lot of regulatory regimes precisely because pre-approving every product before launch would be impossibly slow. The tradeoff is that self-certification is a claim, not a verified fact, until someone checks it. Most of the time nobody checks, or checks so long after the fact that the company has already moved on. Tesla's Cybercab launch is the version of this where someone checked almost immediately, at a scale (roughly 1,000 vehicles, a public paid-rides launch, a lot of press attention) that made the exposure impossible to ignore.
That's the part worth generalizing past robotaxis. If you build software that touches data privacy regulation, accessibility requirements, payment processing rules, or any other area where you're the one attesting "this meets the requirement" rather than getting a third party to verify it, you are making the same bet Tesla made, just with far less scrutiny attached because you're not deploying a fleet of visibly wheel-free cars on a public street. The bet usually pays off, because most solo products never get audited. It stops paying off the moment a customer, an insurer, or a regulator decides to actually look, and by then the gap between "I was confident this was compliant" and "someone verified this is compliant" has usually existed for a while, unexamined.
What I'd do differently
I'm not saying go get formal certification for everything, that's neither realistic nor proportionate for most solo-built products. What I'd actually do: for the two or three places where your product makes a compliance claim you haven't had independently verified, write down what the claim is, what you're basing it on, and how confident you actually are versus how confident you've been assuming you are. If a client, an enterprise deal, or a platform's terms of service ever asks you to substantiate one of those claims, that's the moment self-certification either holds up or doesn't, and I'd rather find the gap on my own schedule than on a regulator's or a lost deal's.
The honest take
I don't think self-certification is inherently reckless, most regulatory systems run on it and most of the time it works fine because verification is expensive and low-probability. Where I'd push back on my own framing: Tesla's situation involves genuine safety stakes and a fleet of a thousand vehicles, which isn't a fair direct comparison to a solo operator's SaaS compliance posture in terms of consequence. But the mechanism, "I attested this meets the bar, nobody's checked yet," is identical regardless of scale, and the Cybercab probe is a useful, visible reminder that "nobody's checked yet" is a temporary state, not a permanent one. Worth an honest gut-check on where your own unverified claims are sitting before something forces the question.
Author
Lukas
@lukcombinator