There is no federal safety standard an automated driving system must meet before it carries passengers on a public road. What exists is an order requiring companies to tell the regulator when their systems crash, and a recall power exercised after the fact. Understanding the difference is most of what there is to understand about this regime.
The Standing General Order
NHTSA issued Standing General Order 2021-01 on 29 June 2021, under its authority at 49 U.S.C. § 30166 to require reports and answers to questions. It has been amended more than once — a substantial revision in April 2023 restructured the reporting, and further changes in 2025 adjusted the scope and cadence, in part narrowing what must be reported for driver-assistance systems.
It applies to two populations, and conflating them is the most common error in reading the published data.
- ADS — automated driving systems, meaning SAE levels 3 through 5, where the system performs the entire dynamic driving task within its operational design domain. Manufacturers and operators of ADS-equipped vehicles are covered.
- Level 2 ADAS — partial driving automation, where combined lane-keeping and speed control operate with a driver responsible at all times. This is a far larger fleet, and its crash counts dominate any naive comparison.
Comparing raw incident counts between the two populations, or between two companies with different fleet sizes and different operational design domains, produces nothing meaningful. The order collects incidents. It does not collect exposure — miles driven, in what conditions — in a form that supports a rate.
This describes a federal reporting order and is not legal or safety advice, and reporting obligations attach to specific entities under specific definitions. Confirm the current text and amendments at NHTSA’s standing general order page.
What triggers a report
The reporting condition turns on engagement in a time window: the system must have been engaged at any point within thirty seconds before the crash. That window is deliberate, and it captures the disengagement case — a system that hands control back immediately before impact still produces a reportable event, which is precisely the scenario that would otherwise disappear from the data.
The order sets tiers by severity and by speed of reporting:
- An incident report within one calendar day of receiving notice, for the most serious categories — a fatality, an injury requiring hospital treatment, a vehicle tow-away, air bag deployment, or involvement of a vulnerable road user such as a pedestrian or cyclist.
- An updated report ten calendar days later, adding what has been learned since.
- Monthly reports for other reportable crashes, including property-damage-only incidents for ADS.
NHTSA publishes the incident data. Two features of the published version limit what can be concluded from it. Fields are redacted where a company asserts confidential business information, and the narrative fields are frequently among them, so the circumstances of an incident are often unavailable. And reporting depends on the entity receiving notice, which favours operators with connected telemetry: a company whose vehicles report crashes automatically will appear in the data more completely than one relying on customer reports, and that is a reporting artefact rather than a safety difference.
No federal agency approves an ADS
Federal Motor Vehicle Safety Standards are set at 49 C.F.R. Part 571 and cover crashworthiness, lighting, braking, electronic stability control and much else. Not one of them specifies how well a vehicle must drive. There is no FMVSS for object detection, for behaviour at an unprotected left turn, or for handling a construction zone.
Compliance is also self-certified. Under 49 U.S.C. §§ 30112 and 30115 a manufacturer certifies that its vehicles conform to the applicable standards; there is no pre-market approval, no type approval, and no federal sign-off of the kind aviation requires. When a company says its vehicle is “fully compliant with all applicable federal motor vehicle safety standards”, that statement is true and says nothing whatsoever about the driving software.
Two adjacent mechanisms are worth knowing. Where a vehicle cannot comply with an existing standard — a design with no steering wheel, for instance — the exemption process at 49 C.F.R. Part 555 allows a limited number of vehicles per manufacturer per year, and the statutory cap is small enough to constrain commercial deployment rather than merely to slow it. Separately, NHTSA amended the 100-series occupant protection standards in 2022 to remove assumptions about manual driving controls, which cleared a definitional obstacle without creating any performance requirement. A national voluntary review and oversight programme for ADS-equipped vehicles has been proposed through rulemaking; a proposal is not a rule, and until it is finalised it imposes nothing.
Defect authority is the real control
The operative federal power over driving software is the defect and recall authority at 49 U.S.C. §§ 30118 to 30120. NHTSA may make an initial decision that a vehicle contains a defect presenting an unreasonable risk to motor vehicle safety and require the manufacturer to remedy it. Nothing in that language requires a broken part; a behaviour can be the defect.
This is why software recalls of driver-assistance systems have become the visible federal action in this space, addressing matters such as the adequacy of driver-monitoring and engagement controls, delivered as over-the-air updates. The Office of Defects Investigation runs a published escalation — a preliminary evaluation, then an engineering analysis, then a recall — and it also opens recall queries to assess whether a remedy already delivered actually worked, which is a distinctively software-era problem: the fix ships in a day and whether it fixed anything is measured over months.
The structural weakness is timing. Defect authority is retrospective and reactive; it acts after a pattern is visible in crash and complaint data. For a system that updates continuously, the population under investigation changes during the investigation. That is a real limitation of the current arrangement and not a criticism of the agency — it is what happens when a statutory scheme designed for mechanical parts is applied to a fleet whose behaviour is redefined overnight.
Where the states pick up
The traditional division is that the federal government regulates the vehicle and the states regulate the driver, registration, insurance and the rules of the road. Automated driving cuts across it, because the driver is the vehicle.
States have filled the space unevenly: some require permits and data reporting before testing or deployment, some have authorised operation by statute with minimal conditions, and some have done nothing. Federal preemption at 49 U.S.C. § 30103(b) prevents a state from establishing a motor vehicle safety standard covering the same aspect of performance as an FMVSS — which, given that no FMVSS addresses driving behaviour, leaves considerable room. The result is a genuine patchwork, and it is being litigated and legislated at both levels. The broader state-level picture is tracked in the state AI law tracker page.
Liability, meanwhile, is not allocated by any of this. When an automated vehicle causes harm, the claim proceeds under ordinary state product liability and negligence law, and the questions — what the design should have been, what warnings were adequate, whether a human was responsible for supervising — are answered by courts case by case. Compliance with the Standing General Order is not a defence to any of them; it is a source of admissions.
Top comments (0)