In January of 2026, Tesla got attention by offering rides in their Austin, Texas robotaxi service with no Tesla employee in the car. Tesla declared operations to be “unsupervised.” The service began with special model Ys and a limited rider list in a very small area, but eventually grew to a larger area for the general public. On Sept 3, the custom-designed Cybercab model was added to the fleet, and since then Tesla has registered 319 Cybercabs in Texas, and reports suggest they have manufactured over 2,300 units so far.

TL;DR Tesla would be crazy to bet the company on wide deployment of an untested vehicle, when every other company in the business remotely supervised their early deployments when they removed all staff from the car. Their current pilot would need fewer staff than their other prior deployments and they put Starlink in every car.

This presents the appearance that Tesla has crossed the most important milestone in self-driving, which is making a car that can drive at a super-human level of safety in a viable service area, without the need for (nearly) full-time human involvement. This is the key bar both because it’s hard, but also because if you need a full time human for each car, it doesn’t scale You might as well just run a taxi service with human drivers. (There is a counter to this, discussed below.)

But has Tesla actually crossed that milestone? It’s likely the answer is no, and that all Tesla robotaxis have a full-time or nearly full-time employee watching the vehicle remotely, able to hit an emergency stop button, and probably more than that. While removing the need for an employee in the vehicle, which they have done, is a big and important step, it’s much smaller than the really big step, not needing a human anywhere full time--that is the step that truly matters, and makes the difference between having an autonomous, unsupervised vehicle, a robotaxi, and not having one. It’s hard to overstate just how big a step that is, so it’s important to learn if Tesla has truly done it, or is just attempting to give the appearance of doing it. Ironically, Tesla’s use of the word “unsupervised” on day one is actually evidence that it’s not unsupervised, not just because they are “The boy who cried ‘self-driving.’”

Absolute Everybody Does It

Having remote supervision is no shame. In fact, everybody starts that way. Waymo, Cruise, Zoox, May (unconfirmed), Nuro, Aurora, all the Chinese companies--anybody who has put a real robotaxi on the road has does this because it’s the only sane thing to do. To not do it would be extraordinary and reckless. After working for years to get a car to the point that you can take out the employee, no company is going to just wave goodbye to the vehicle, and send it off into the world empty and not watch, and not be ready, just in case, to make it stop if something goes wrong. It’s not expensive with a small volume of vehicles, and because one mistake could kill your project, it’s worth almost any cost.

Several of these other companies, however, later confirmed they stopped doing this. They rarely announce it, but have disclosed it when pressed. The removal of the safety driver looks like the big step and gets the most press, even though it’s a big step and not the big step. Waymo, Zoox and Aurora’s executives have confirmed to me directly that they no longer remote supervise full-time.

Tesla have not stated they have stopped remote supervision. Alas, Tesla no longer responds to queries from the press; they take questions only when they feel like it, mostly on X, and from a select group of analysts on earnings calls. Probably all Tesla has to do is just make a much more explicit declaration and that will big a big deal. But they don’t. The correct presumption for all companies, not just Tesla, is that they are using remote supervision until they explicitly declare they have stopped doing it.

Realize that Tesla is claiming something extraordinary here, that nobody has ever one--a sufficiently safe self driving system that needs no supervision based just on cameras and end-to-end neural nets. A thing many (but far from all, and not including myself) experts have opined is impossible. They don’t agree it’s impossible, but all agree it’s never been done before, so the presumption is that it hasn’t until clear proof is presented.

Remote Supervision vs. Remote Assist vs. Remote Driving

These three terms cause some confusion. They are three pretty different things, though they all involve a remote employee at a console and a data connection.

  1. Remote Supervision is a person remotely watching the vehicle do the driving. That person has some special commands, like “emergency stop” and “pull over.” These are commands Tesla’s right-side safety drivers had (Tesla called them safety monitors) though they also had, like a driving instructor, the ability to grab the wheel and steer the car. Remote supervision wants a decent data connection, but it does not demand perfection. Every company uses remote supervision at first, with plans to discontinue it when they can. A remote supervisor can typically command that the car enter remote assist mode, and they can be the assist operator. They can probably in Tesla’s case, also command going to remote driving mode if needed. It is speculated some companies have one remote supervisor watch multiple cars; if so, it’s not very many.
  2. Remote Assist does not have a human watching full time. Rather, the cars decide if a situation is too confusing and they request help from a human. The car is almost always stopped at that point, and the car offers the human a menu of options the car generates. The human can also make up their own solution but this is rare. In fact, most of the time, the remote assist operator just confirms the car’s first choice. Remote assist can be done over a poorer, sometimes high-latency data connection from anywhere in the world. Waymo famously suffered when too many of their cars demanded remote assist advice at once during a power failure that turned off traffic signals, overloading their assist center. Cars waited for long periods listening to virtual hold music. Every robotaxi company uses remote assist, including Tesla and Waymo, and plans to do so effectively forever. Waymo reported they have about one remote assist operator for every 30 cars. Once you get to 50 cars or more, there are diminishing returns, and you are ready to scale up fine without high labor costs.
  3. Remote Driving is done by a human using a remote, video-game style wheel and pedals and a bunch of screens or VR goggles. The human makes decisions on direction and speed in near real time. The car still has self-driving ability, and it handles low end tactical moves, and it handles them entirely if the latency gets too large or the connection goes out. If it goes out for too long, the system usually pulls the car over. This requires a decent connection, but surprisingly not as good as you probably think. It’s also possible to do full tele-driving, more like an RC car. That requires a very high quality data connection. So high quality that effectively nobody does this except at short range. Tesla has a remote driving system. They are the only robocar company I know of to make one. They have stated it is only activated at very low speeds (below 2mph) and only drives up to 10mph, but others have alleged based on rides that this may not be true.

Evidence Tesla Still Uses Remote Supervision

Since everybody starts with remote supervision, you don’t need evidence for that. It’s just assumed. It would be reckless to start without it. The real question is when somebody stops remote supervision.

Tesla has called their service unsupervised. Unfortunately, they did this from the very start of the service. As such, it’s extremely likely they are using the word “unsupervised” to mean “not supervised inside the car.” Sadly, there is a long history, both with Tesla and with many other companies in the space to exaggerate the level of autonomy in their system. I have seen large numbers of announcements from companies declaring they now have “autonomous” cars or “robotaxi” service when there is still a safety driver on board. Quite often they have released video that makes it hard to see the safety driver unless you look carefully. They may not mention them in the release, or mention it in very fine print.

We’ve also seen multiple companies, including Tesla and AVRide, make a big deal out of moving the safety driver from the left seat to the right seat. “Nobody behind the wheel.” As though this makes a difference. A billion people have learned to drive with their driving instructor on the right side able to grab the wheel and with an emergency stop brake. It works. It had better work--if moving the safety driver to the right side is actually dangerous, they should not be doing a stunt like this that puts people at more risk, since there is no business or technological reason to do it. It’s just PR.

Tesla, sadly has a long history of fudging the meaning of words. When they called their driver-assist tool “full self driving” it raised an uproar, because it obviously was not, and both California and Europe had regulators actually force Tesla to stop using that name. As such, when Tesla says “unsupervised” the priors suggest we be highly skeptical.

  1. Tesla has put special communications gear in each robotaxi. Cybercabs come standard with Starlink. The model Ys have a special communications box in the rear which may contain Starlink and mobile data antennas. Starlink has very low latency but does not have 100% coverage. The combination of Starlink and 5G networks combined offers very good service. Tesla’s special relationship with SpaceX (the CEOs are good friends) could enable Tesla getting special priority and bandwidth from Starlink. (Normally, Starlink upstream is lower than downstream.)
  2. Tesla is known to have remote driving. They hired developers to build it and there are photos of remote driving consoles in their HQ. Tesla admits they have it but claims it is used only at low speed. This means they definitely have the ability to watch through the vehicle cameras. Indeed, any private Tesla offers the owner the ability to look out the cameras in real time over 5G networks.
  3. Tesla has a disappointing history of “fake it till you make it” operations, including demonstrations of the Optimus robot that were revealed only later to be just remote controlled. Tesla has a long history of publishing safety statistics on Autopilot and FSD (Supervised) in a highly misleading form, using words in nonstandard ways to look better than they are. Sadly, that makes us skeptical of their claims and demos.
  4. Tesla’s roll-out of vehicles and Cybercabs has been slow, but speeding up. While there are many factors which might slow a roll-out, the need to have human supervision is a compelling one. Elon Musk has regularly promised the service will scale extremely fast. To not deliver on this requires a very compelling reason. Reports suggest teams that disappoint the CEO face severe consequences.
  5. Musk and his team have repeatedly indicated they know how important it is they don’t have a serious incident in their project’s early history. They have seen how Cruise and Uber ATG were shut down, wasting billions, when they failed that test. No matter how pleased you are with your vehicle’s performance, when you only have a few million miles of vacant operation under your belt, you can never be sure enough to bet your whole project on it not making a mistake. That’s what human supervision helps prevent, and why everybody uses until they get the real data needed to justify removing it.
  6. At the Cybercab roll-out they reported one million miles of operation for their robotaxis. Tesla’s crash reports, which largely predate the Cybercab roll-out, show 10 at-fault crashes (generally minor) and 14 not-at-fault, as well as 4 that were the fault of a remote driver. 10 crashes in one million miles is a lower-than-human crash rate, and generally a team would not stop remote supervision until much better than that rate. Waymo reports one liability incident (a subset of at-fault crashes) every 2.3 million miles. Tesla might be reckless, however. 14 crashes (including those caused by remote driver) is a quite poor rate for a supervised system. Tesla has shown that their consumer FSD supervised product has a much better rate than that (though that rate is for major airbag-deployed crashes.) Remote supervision is not as good as in-car supervision, however. Tesla’s rate of 14 not-at-fault crashes is actually higher than average: Just driving around you will get hit a certain amount, it is mostly beyond your control. The reason for this is unclear.

Can you do remote driving and supervision over an unreliable data network? Yes, you can, and several have done it. Now it turns out the combination of Starlink and multiple 5G networks can produce a very reliable, almost always low-latency data connection, even though Starlink has blackouts when the sky is blocked or the satellites have a gap. T-mobile in Austin is often under 20ms. Starlink is about 35ms. Average human driver reaction times are more like 300-700ms. (Drunks are 1,500 to 2,000ms.)

A remote supervisor doesn’t drive, though, they just think ahead and recognize dangerous situations where an emergency stop or other interventions are called for. Say you have a car that has one incident every 10,000 miles, as Musk claimed Tesla had in 2025. Even if your supervisors can catch only 9 out of 10, due to their own mistakes or data outages, you still will improve that to 100,000 miles, which is not fantastic but approaching the threshold of acceptable. If the car has improved to say every 30,000 miles, that is still way to much to deploy unsupervised, but the 300,000 miles you could get supervised is entirely reasonable. By limited the roads the car drives to a limited area and staying away from dangerous streets, and not doing freeways (as Tesla robotaxi does not support) you can easily meet targets, even with a poor car and a data network with regular outages. Network outages are independent of car errors, and outages of cell networks are not correlated with Starlink outages.

So what does Tesla have to do to establish they have really built an unsupervised, safe robocar with cameras and E2E neural nets? The first thing would be to get explicit in their declaration, ideally in a communication to shareholders. No equivocal use of words. Better would be to let an external 3rd party come in and confirm it. Absent that, they could deploy a system so big it is unlikely to be supervised--perhaps 5-10 million miles/week, though it’s not clear if a “bet the company” move doesn’t justify almost any budget.

It’s a Great Strategy for Tesla

Doing a remote supervised robotaxi service with up to 2,000 cars is actually a great strategy for Tesla to get real-world experience driving roads empty and carrying passengers. This experience is vital to get towards deploying. Tesla operates a service in California with human safety drivers in each car, with over 700 cars registered for it. Remote supervisors take only 1/3rd to 1/2 the labor of in-car staff, because they are not needed when the car is parked waiting for a rider, charging, or even stopped at a red light. Since Tesla is willing to hire staff to sit in 700 California cars, they are obviously ready to do the same for about 2,000 remotely supervised cars. That is, this is easy if you are one of those who just happens to also run a space internet company that you can get cheap/free spare bandwidth from. As so many people do.

Tesla’s most recent announcement was they have just now reached an impressive 428,000 miles per week. At the average Austin speed (not counting red lights) of 21mph, that’s about 3,000 hours/day of supervising. At 7.5 hours/day that’s under 400 shifts per day of supervising. That’s fewer people than they have as they have in California, where if the 759 cars there are doing 2 shifts/day is 1,400 shifts. (The cars probably don’t drive all day.) The old Model Y fleet in Austin had 133 cars, back in August, so about 260 shifts/day.

The reality that Waymo learned as they deployed is there’s a whole ton of surprises that cause problems that you never see until you take out the employee. With a large remotely supervised fleet, Tesla can learn those things even before they are ready to go unsupervised.

The Boy Who Cried “Self-Driving”

In the fable of the “Boy who cried wolf" we know that after many false claims, a real wolf appeared and nobody believed the boy any more. But the villagers weren’t being stupid, they were making the right decision based on their prior experience. It was the boy who made the mistake with the false alarms. Tesla, and Musk, are the boy who declared “self driving” every year for over 10 years. Should the day come when they actually have real, full, self-driving, smart people won’t believe them. They have earned skepticism, and the duty is on them to make an extra effort to show it’s real. In spite of this, they remain opaque, using ambiguous words, redacting reports, and releasing only limited data. If people disbelieve then when they actually have a wolf, the fault will lie with them.

Since Tesla doesn’t give us much information, and they have cried wolf dozens of times, how can one decide what to believe. Compare these two stories:

  1. Tesla decided to send vehicles out with no employee. They fired or transferred the 260+ in-car “Safety Monitors” and in spite having very few test miles in a vacant vehicle (but plenty in supervised vehicles) they sent out a decent sized fleet of empty vehicles, including the brand new, barely tested Cybercabs with nobody in them. Unlike every other self-driving company, they simply sent the vehicles out,hoping for their safe return, without monitoring. Even though everybody else felt they needed millions of miles of empty vehicle operations with monitoring before doing such a thing, and even though Musk and other Tesla executives have declared they are betting the $1.5T valuation company on robotaxis and Optimus, and even though the declared it was essential they not make a big mistake, having seen how Cruise and Uber ATG were shut down after such mistakes, they decided to throw caution to the winds and deploy and grow fast just to save some money from firing those staff. Or…
  2. Like Waymo, Cruise, Zoox, Nuro and the Chinese companies, they arranged so a human could watch the newly empty vehicles over the network. They used the existing staff to watch and gave them an emergency stop button like what they had when they sat in the car, plus a remote control wheel which they rarely used when they were in the car. They installed multiple 5G radios, better antennas and Starlink that they installed in every car, thanks to a special deal with SpaceX only Elon Musk could arrange. They could now quickly, and with confidence, deploy and grow the deployments and the service areas., even in to funny shapes. They are learning, and gathering data until they finally have the safety data to prove they can safely deploy with nobody in the car, and rare remote monitoring. (Remote assist is there as always for all projects.) Rather than bet the company on something they had never done before, they took the safe path with a human fallback in case they, like every other company out there, discovered they made mistakes and had things to learn you only learn when you go out empty. They’re getting tons of useful experience and data. Because they have bet the company there is almost no limit on how many cars they might deploy this way, other than there are diminishing returns if you start doing many millions of miles per week. Like the other companies, they didn’t make a PR announcement about the remote supervisors--those other companies only revealed they had stopped it several months later, when pressed.

Which scenario makes the most sense? Which would you do if you were in charge, and knew a serious error by your untested vehicle could throw away your whole project, your whole company, and you have warned about that.