A car can steer, brake and hold its speed while you sit there doing almost nothing.
You're still driving.
Another car can carry a passenger across town with an empty driver's seat. It may be driving, but only inside a mapped area under conditions its operator has approved.
People call both cars "self-driving," leaving you to work out what each one requires of you.
SAE International has six levels for this stuff. The levels describe who is responsible for the driving task.
Level 0: The car warns you. A blind-spot warning flashes. Automatic emergency braking may jump in for a moment. You remain responsible for driving. The risk is small but familiar: a warning becomes background noise, or a driver assumes the safety feature can see more than it can.
Level 1: The car handles one control. Adaptive cruise control can manage speed. Lane centering can steer. You watch the road and handle everything else, controlling speed and braking while the system steers, or steering while it controls speed.
Level 2: It feels like the car has the job. Now the system can steer and control speed at the same time. This is where people get comfortable.
SAE and NHTSA still put the driving job on the person. You keep your eyes on the road, ready to brake or steer, even after thirty miles without a problem.
The National Transportation Safety Board has spent years pointing at the problem. Humans are lousy at watching a machine perform correctly for long stretches and then snapping back to full attention for the rare failure.
In March 2018, a Tesla Model X operating with partial automation steered into a highway gore area in Mountain View, California. The driver was distracted and didn't respond. The SUV hit a damaged crash barrier at about 71 mph. The driver died.
The NTSB blamed more than one failure: limits in the system, the driver's lack of response, ineffective driver monitoring and a barrier that hadn't been repaired.
The driver needed to remain attentive despite the automation.
Level 3: The system may ask you to take over. At Level 3, the system handles the driving task under defined conditions. It asks the person to take over when it can no longer continue.
That handoff is harder than it sounds. The person has to look up, work out where the car is, understand what changed, find the hazard and make the right move. The warning has to allow time for that.
A late warning may leave the person too little time to take control.
Level 4: The machine drives inside a fence. At Level 4, passengers aren't expected to take over while the system operates. The catch is the fence — a city, certain roads, approved weather, a speed range, or some combination. A driverless taxi in a defined service area is the usual example.
The vehicle may own the driving task. The company still needs people watching the fleet, handling collisions, working with police and shutting down service when the operating conditions go sideways.
Level 5: Everywhere a person can drive. Level 5 means the machine handles the entire driving task under every condition and on every road where a person could drive. The definition covers driving without a mapped service-area limit, weather exception or takeover request.
That's the version people picture when they hear "self-driving car." It isn't what today's consumer systems deliver. NHTSA avoids the loose phrase because the name can make drivers assume a capability the vehicle doesn't have.
Read the manual to find out what the system requires of you.
What this teaches us about AI
The SAE levels apply to vehicles. A chatbot doesn't earn "Level 4" because somebody put an agent label on the box. You still need to ask who is responsible for the action.
Is the tool giving you advice? Writing a draft? Preparing an action for approval? Taking a narrow action within limits? What happens when it reaches something outside those limits?
Those are different products. They need different controls.
A chatbot rewriting a paragraph can run with plenty of freedom. An agent sending the paragraph to 4,000 customers should stop and show you the final copy, recipient list and account it will use.
Set controls according to what the action could affect, including whether you are deleting duplicate photos or a client's project folder.
Give one task a level
Pick one thing you want AI to do. Write the task at the top of a page. Then choose the control:
- It gives advice. You do the work.
- It prepares a draft, and you submit it.
- It shows you the exact action, then waits for approval.
- It acts by itself inside written limits and stops at the edge.
Underneath, write three facts: what it may touch, how much it may spend and what it must never do without you.
Those details give you a rule you can check before the tool acts.
One question back
Name one task and choose one control from the list above.
Keep client names, credentials and private documents out of the reply.
Sources
- SAE International, J3016_202609: Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles (current revision listed September 20, 2026).
- National Highway Traffic Safety Administration, Automated Vehicle Safety.
- National Transportation Safety Board, Vehicle Automation: Investigative Outcomes and Recommendations.
- National Transportation Safety Board, Mountain View, California, partial automation crash investigation.
Brad Heath and Josh are co-founders of Old Guyz AI. Brad is its in-house software developer.