As at 3 November 2026, an Australian driver using Tesla Full Self-Driving (Supervised) remains responsible for operating the vehicle. FSD is a supervised driver-assistance system, not an autonomous driver. But that does not mean the human must necessarily bear sole responsibility for every accident involving it. If the system itself contributes to a collision, questions of causation, product liability, negligence and the driver’s actual opportunity to intervene may arise. Australia’s proposed automated vehicle laws will eventually go further by placing responsibility for genuine automated driving systems on a corporate Automated Driving System Entity.
On 14 September 2026, the gap between those two worlds became harder to ignore.
Zurich announced that Tesla Full Self-Driving (Supervised) would become a rating factor for its InsureMyTesla product. Zurich described the change as a first for the Australian insurance industry and the second instance of an insurer doing so globally. It relied on Tesla’s US data, which reports seven times fewer major or minor collisions for vehicles using FSD than the estimated average for all vehicles, and said local Tesla customers had already driven more than 132 million kilometres using the technology.
The development is interesting for a reason that has little to do with the size of anyone’s insurance premium.
An Australian insurer is now prepared to distinguish the risk associated with machine-assisted driving from ordinary driving. Australian road law, for the moment, largely continues to identify the human behind the wheel as the driver.
Those propositions can coexist while nothing goes wrong.
A serious crash makes the distinction much harder.
Suppose a Tesla is travelling through Adelaide with FSD (Supervised) engaged. The vehicle identifies the wrong path through an intersection, steers across the path of another vehicle and causes a collision.
The person behind the wheel says they were looking at the road. They did not command the steering movement. By the time they realised what the car was doing, the collision could no longer be avoided.
Who was driving?
The present legal answer is the human.
The harder question is whether that tells us who actually caused the dangerous movement, whether the human had a realistic opportunity to stop it and, ultimately, who should bear responsibility for what occurred.
That is the problem automated driving creates for a body of road law built around a much simpler proposition: the driver controls the car.
Who is liable if Tesla FSD causes a crash in Australia?
Tesla itself does not describe the Australian version of FSD as autonomous driving.
FSD (Supervised) can change lanes, select forks to follow a navigation route, navigate around other vehicles and objects and make left and right turns. Tesla nevertheless states that none of those functions make the vehicle autonomous or replace the driver. The driver must remain attentive and maintain proper control.
Australian regulators presently draw the same line.
The National Transport Commission says Australia’s current road laws apply to vehicles fitted with advanced driver assistance systems, including lane keeping and adaptive cruise control, with a human driver remaining in control. Its automated vehicle program is instead directed to SAE Levels 3 to 5, where an automated driving system performs the entire dynamic driving task within its operational design domain. Those vehicles are not yet commercially available for general use on Australian public roads.
For a South Australian driver using FSD (Supervised) today, “the car was driving itself” is therefore not, by itself, an answer to a traffic charge.
But neither should the analysis end with the label “driver”.
The real legal question is what the human did, what the system did and whether the conduct required to establish the particular offence can actually be attributed to the accused.
What South Australian law says about driver responsibility
South Australian traffic offences were drafted on the assumption that a human driver’s decisions and the physical movement of the vehicle are ordinarily connected.
Section 45 of the Road Traffic Act 1961 (SA) prohibits driving without due care or attention or without reasonable consideration for other persons using the road. Talon Legal’s detailed guide to driving without due care in South Australia considers the ordinary operation of that offence.
Section 46 prohibits driving recklessly or at a speed or in a manner dangerous to any person. The distinction between careless and dangerous driving is considered further in our guide to dangerous driving in South Australia.
Where death or injury results, the Criminal Law Consolidation Act 1935 (SA) becomes particularly important. Section 19A deals with causing death or harm by dangerous use of a vehicle or vessel. Section 19ABA separately creates offences where a person drives without due care or attention, or without reasonable consideration for any person, and by that conduct causes death or serious harm.
The language matters.
The legislation does not simply punish being present in the driver’s seat when something bad happens. It identifies a person who drives in a particular way and, for the serious-result offences, connects the prohibited driving conduct with the resulting harm.
That relationship is obvious enough where a human sees a gap, turns the steering wheel and collides with another vehicle.
It is less obvious where software sees the gap, selects a trajectory and issues the steering command.
The prosecution case may then have to be framed differently. The alleged fault is no longer necessarily that the accused positively steered into danger. It may be that the accused remained legally responsible for supervising the automated system and failed to intervene when a reasonably attentive driver should have done so.
That is not merely a semantic difference.
The criminal allegation has moved from performing the dangerous act to failing to prevent the machine from performing it.
South Australian law already recognises cars which override their drivers
There is a useful clue in the Road Traffic Act itself.
Section 44C prohibits a person from driving an ultra high powered vehicle on a road if an automated intervention system of the vehicle is disabled. It defines an automated intervention system as one that identifies a hazardous condition and responds by overriding the driver’s control of the vehicle, such as anti-lock braking, automated emergency braking, electronic stability control or traction control. The legislation therefore already accommodates vehicles in which electronic systems intervene in the driving process rather than merely report information to the human.
That arrangement causes little conceptual difficulty with anti-lock braking or electronic stability control. The human remains plainly responsible for driving and a safety system intervenes at a particular moment.
FSD and comparable systems reverse the balance.
The software may continuously perceive the road, choose a path, regulate speed and steer the vehicle, while the human’s practical role becomes supervision and occasional intervention.
The law still calls the human the driver. The engineering reality is beginning to look different.
What if the car, not the driver, makes the dangerous move?
Consider two accidents involving the same vehicle and the same software.
In the first, FSD begins moving towards another lane. The danger develops over several seconds. The driver is looking at a phone, ignores warnings and makes no attempt to intervene.
In the second, the driver is watching the road. The system suddenly commands an unsafe steering movement and, by the time the error becomes perceptible, the collision cannot reasonably be avoided.
Both drivers have a steering wheel in front of them.
Both can theoretically touch the brake.
It does not follow that both had the same control over the event.
The more useful question is whether the person had meaningful control when it mattered.
That concept is not presently a statutory test. Nor has an Australian appellate court established a doctrine of “meaningful control” for automated vehicles.
It is, however, the issue hidden inside the facts.
Where criminal responsibility is said to arise because a human failed to override automation, a court should ultimately need to know whether there was a reasonable and effective opportunity to do so.
The intervention window
That opportunity can be described as the intervention window.
It is the period between the point at which the developing danger ought reasonably to have become apparent to a properly attentive driver and the last point at which human intervention could still avoid or materially reduce the collision.
The difference between 300 milliseconds and eight seconds is enormous.
If a vehicle spends eight seconds drifting towards danger while the driver looks away, the existence of automation may do little to assist the defence.
If the software makes an unexpected decision and the collision becomes inevitable before a human could reasonably identify and reverse it, saying that the accused “could have grabbed the wheel” proves very little.
This distinction already has an analogue in conventional traffic law. Our guide to driving without due care discusses the latent defect defence, where a driver loses control because of a mechanical defect without any fault of their own.
Automated driving makes the problem harder. A software-controlled vehicle need not be mechanically defective at all. The steering actuator may work perfectly. The cameras may work. The computer may simply make the wrong driving decision.
The failure is no longer necessarily in the machinery.
It may be in perception, classification, prediction or planning.
That is why future automated-driving cases are likely to require a different kind of accident reconstruction.
The Australian road environment is already exposing the gap
In July 2026, University of Queensland researchers reported on more than 100 days of driving a Tesla Model Y using FSD.
They recorded more than 500 safety-critical events. They reported recurring difficulties with school zones, railway crossings and boom gates, complex roundabouts, zipper merges, road markings and harsh or variable weather. They did not complete a single trip in which FSD drove from start to finish without intervention.
Their observations were not a simple finding that the system was unsafe. The researchers also found that in many situations it drove with a precision a human driver would find hard to match.
That combination is the interesting part.
A system can become good enough that a human increasingly trusts it without becoming good enough that the human can safely stop supervising it.
Present Australian law depends heavily on that distinction. The human must remain attentive precisely because the technology can occasionally make decisions which require immediate correction.
That creates an unusual human-factors problem. The better the machine becomes at doing the routine work, the more unnatural continuous vigilance may become.
A warning telling the driver to remain attentive is plainly relevant to that person’s responsibility.
It does not necessarily answer whether the human-machine system was reasonably designed in the first place.
Can Tesla or another manufacturer also be liable?
Driver responsibility and manufacturer responsibility are not mutually exclusive.
Australian product liability law already provides a framework for claims involving goods with safety defects. Part 3-5 of the Australian Consumer Law (liability of manufacturers for goods with safety defects) can make a manufacturer liable for loss or injury caused by goods with a safety defect, in the circumstances that Part sets out. Ordinary negligence principles may also arise depending upon the facts.
An automated-driving collision could therefore involve several separate questions.
Did the driver supervise the system as required?
Did the system correctly detect the hazard?
Was the vehicle permitted to use automation in circumstances where its capabilities were inadequate?
Did driver-monitoring systems respond appropriately?
Were warnings adequate?
Was a sensor incorrectly calibrated?
Did a software update alter the system’s behaviour?
The answer does not have to be that one side was entirely responsible and the other entirely blameless.
American litigation already provides an example.
In Benavides v Tesla Inc, a federal jury in Florida considering a fatal 2019 crash involving Autopilot allocated 67 per cent of fault to the driver and 33 per cent to Tesla in August 2025. Tesla’s own securities filings record the verdict and the allocation of responsibility, and Tesla has appealed. The case remains American product-liability litigation and does not determine Australian law, but its significance is obvious: human inattention and alleged defects in an automated system can be treated as concurrent causes of the same accident.
That distinction will matter even more in criminal proceedings. Civil law can divide responsibility between several causes. Criminal liability remains concerned with whether the conduct and culpability of the particular accused satisfy the offence charged.
What Australian courts have said so far
Australian authority directly addressing this problem remains sparse.
The most useful reported example is DPP v Agrawal, a 2024 sentencing in the County Court of Victoria.
The case involved a Tesla Model 3 which struck a pedestrian who was boarding a tram. The driver had claimed the car was on Autopilot, but the car’s data showed Autopilot was not engaged before the collision. The driver pleaded guilty to dangerous driving and failing to stop.
Agrawal is sometimes tempting to describe as an “Autopilot case”. That would overstate what it decided.
The Court did not determine who bears legal responsibility where Autopilot itself causes the accident-producing movement.
The important point is evidentiary.
The car was capable of contradicting the driver’s account.
No reported Australian decision located in preparing this article squarely determines the more difficult case: a Level 2 system is genuinely operating, the system commands the dangerous manoeuvre, the human remains legally required to supervise it, and the issue becomes whether the human had a sufficient opportunity to prevent what occurred.
That case is still coming.
Tesla data could become crucial evidence after a crash
A conventional traffic investigation may involve witness accounts, CCTV, skid marks, damage patterns, speed calculations and vehicle inspections.
A serious FSD accident potentially adds an entirely different layer.
The vehicle may be able to reveal whether the automated system was engaged, when it disengaged, vehicle speed, steering inputs, braking, warnings and other operational information.
More sophisticated litigation may need to go further and reconstruct the automated decision itself.
What did the cameras perceive?
How did the system classify the object?
What did it predict that object would do?
What path did it select?
What steering or braking command followed?
What warning did the driver receive?
When did the human intervene?
Tesla’s own safety methodology demonstrates why seemingly minor timing questions will matter. For its safety reporting, Tesla classifies a collision as occurring with FSD engaged if FSD was active at any point in the five seconds before the collision, including where the driver disengaged it or the system aborted shortly before impact. Tesla also says it does not attribute fault in its collision reporting.
A five-second statistical rule may be perfectly sensible for fleet safety analysis.
It is not a rule of legal causation.
A court deciding a particular case will want to know what happened during those five seconds.
That is why preservation of vehicle and software evidence may become critical from the outset of a serious matter. A modern vehicle can also change after an accident through over-the-air software updates. Tesla says FSD will continue to improve with future software updates.
The car inspected months later may therefore be physically the same vehicle but not necessarily running the same software that controlled it on the day of the collision.
In some future cases, preserving the version of the car may matter almost as much as preserving the car itself.
Why Zurich’s Tesla FSD insurance decision matters
Zurich’s September 2026 decision shows that the insurance market is already beginning to distinguish human and automated driving risk.
But the detail is more interesting than the headline.
According to CarExpert, Zurich treats an active FSD (Supervised) subscription as the rating factor, and Insurance Business reported that Zurich would not collect individual driver or vehicle telemetry from Tesla. The premium therefore turns on having FSD, not on whether it was engaged in any particular crash.
That separates three questions which are often blurred together.
An insurer may ask whether people who subscribe to FSD represent a different actuarial risk.
A civil court may ask what actually caused a particular collision.
A criminal court may ask whether the conduct of the accused driver satisfies an offence beyond reasonable doubt.
They are not the same inquiry.
Zurich cites Tesla data reporting seven times fewer collisions for vehicles using FSD. Tesla’s published methodology, however, compares its FSD collision results with an estimated US average, applies its five-second engagement rule and says it does not attribute fault. The statistic is relevant to risk. It cannot determine liability in an individual accident.
That distinction will become increasingly important.
Insurance is beginning to price the machine.
The law is still deciding when to blame it.
Australia’s Automated Vehicle Safety Law will change the question
Australia is already developing the legal framework for the point at which genuine automated driving arrives.
On 21 November 2025, Australia’s infrastructure and transport ministers agreed a new approach that will enable conditional deployment of automated vehicles in select locations from 2027 as a first step, subject to each jurisdiction’s capabilities.
At the centre of the proposed framework is the Automated Vehicle Safety Law, or AVSL.
The Commonwealth’s present description is unusually clear. The AVSL will place responsibility for the safety of an automated driving system on a corporation rather than the human driver. That corporation will be known as the Automated Driving System Entity, or ADSE, and will carry safety obligations over the system’s on-road life. As at 3 November 2026 the Commonwealth described the AVSL as still being developed.
That is more than another vehicle-safety regulation.
It changes the legal architecture.
Current road law begins with a driver and asks whether that person drove lawfully.
The future framework begins to ask which actor had responsibility for the particular driving function being performed.
The Level 3 problem: when does responsibility return to the human?
Level 4 automation is comparatively easy to understand. Within its authorised operating environment, the system performs the driving task and does not expect a passenger to rescue it.
Level 3 is much harder.
A Level 3 automated driving system performs the entire driving task within its operational design domain, but a human fallback-ready user must be receptive to the system’s request to take over, known as a transition demand.
Australian transport ministers have already decided some obligations for Level 3 users, including a requirement to respond to a transition demand and to hold an appropriate driver licence.
Now imagine the system is travelling at 100 km/h.
It encounters a situation it cannot manage.
A warning tells the user to take control.
Five seconds later there is a crash.
When does responsibility move back to the human?
At the instant the warning appears?
When the human perceives it?
After enough time to regain awareness of surrounding traffic?
After enough time to understand why the system is surrendering control?
Or only when there has been a reasonable opportunity to physically assume control and avoid the danger?
The United Kingdom’s Automated Vehicles Act 2024 addresses precisely this boundary. Section 7 requires the transition period to be long enough for the user-in-charge to prepare to assume, and assume, control of the vehicle, and requires the vehicle to keep travelling safely and legally during it. Under sections 47 and 48, the user-in-charge does not commit an offence arising from the way the vehicle is driven, but that protection ends once a transition demand has been issued and the transition period has ended. The official UK legislation website shows these provisions as not yet in force.
That is the intervention-window problem translated into legislation.
Who should be responsible when the machine is actually driving?
For the moment, a South Australian using FSD (Supervised), adaptive cruise control or lane assistance remains the driver.
That proposition should not be stretched beyond what it actually establishes.
Being the driver does not automatically prove that every physical movement of the vehicle was commanded by the human.
It does not establish that the human had enough time to prevent a particular automated action.
It does not exclude a defect in the vehicle or software.
And it does not remove the prosecution’s obligation to prove the conduct required by the offence charged.
As automation increases, two propositions should remain aligned.
Responsibility should not transfer faster than control.
A driver who ignores warnings, ceases supervising a Level 2 system or deliberately overrides automation should not be able to point at the software simply because it had been engaged.
But the converse is just as important.
Responsibility should not remain with the human after meaningful control has genuinely transferred to the machine.
A brake pedal within reach is not meaningful control if there is no reasonable time to perceive the problem and use it.
That distinction is particularly important in criminal law.
Civil law can recognise concurrent causes. Insurers can price risk across millions of kilometres. Regulators can impose duties on manufacturers and system operators.
A criminal court is deciding something narrower and more serious: whether this accused person committed this offence.
When automated systems begin making the decisions that produce collisions, the traditional question, “Who was sitting in the driver’s seat?”, will increasingly be insufficient.
The better question will be:
Who, or what, had meaningful control of the driving task when it mattered?
That question is likely to sit at the centre of the next generation of traffic litigation.
Frequently asked questions
Is Tesla Full Self-Driving legal in Australia?
Tesla FSD (Supervised) is available in Australia as an advanced driver-assistance system. Tesla expressly states that it does not make the vehicle autonomous and requires an attentive driver who remains responsible for controlling the vehicle. Genuine SAE Level 3 to 5 automated vehicles are being addressed through a separate Australian regulatory framework.
Who is liable if Tesla FSD causes an accident in Australia?
For present Level 2 systems such as FSD (Supervised), Australian road law continues to treat the human as the driver. That does not necessarily exclude separate civil responsibility involving a manufacturer or defective product. In a criminal prosecution, the precise issue will depend upon the offence charged, what the automated system did, what the human did or failed to do and whether the prosecution can prove the required conduct and causation.
Can Tesla data show whether FSD was engaged during an accident?
Potentially. Modern Teslas generate extensive operational information, and DPP v Agrawal demonstrates that vehicle data can be used to test a driver’s account of which automated functions were operating. Where criminal charges or serious injury are involved, preservation of vehicle, diagnostic and software information may therefore become important at an early stage.
Law current at 3 November 2026.
Talon Legal acts in South Australian traffic and criminal matters, including careless and dangerous driving prosecutions. For further information, see our Traffic Lawyers Adelaide practice area.
