Tesla Claims FSD Can Predict Bad Driving After Analysing Millions of Real-World Merges

Tesla’s official account on X posted this week that its Full Self-Driving system, after “watching humans handle merges and exits millions of times”, can “quickly predict and respond to bad driving, often before it happens.” The claim promotes FSD’s machine-learning capabilities and positions the software as anticipatory rather than merely reactive. But the post has drawn renewed scrutiny from safety researchers, independent testers, and former Tesla employees who question whether the promotional language reflects the system’s real-world performance.

FSD is, in legal terms, a Level 2 advanced driver-assistance system. That classification — used by regulators including the US National Highway Traffic Safety Administration and reflected in UK and EU law — means a human driver must remain fully attentive and responsible at all times. It is not self-driving. It is not autonomous. The name, critics argue, implies something the technology is not.

How Tesla Trains FSD

The tweet’s reference to “watching” millions of merges and exits describes a real process. Tesla collects camera and sensor recordings from its global customer fleet and uses neural networks to identify patterns in human driving behaviour — including how drivers handle motorway merges, slip roads, and complex exits. Human “data labellers” annotate examples of good and poor driving, and the system is trained to recognise and respond to both.

Tesla calls this “fleet learning.” According to the company’s own AI Day and Autonomy Day presentations, the scale of data involved is genuinely large, and the approach allows FSD to anticipate hazards — a vehicle braking sharply at an exit, say, or a driver cutting across lanes — and respond before a human might react. That is the capability the tweet is promoting.

But the quality of that training data has been questioned by people who did the labelling work. Former Tesla data labellers have publicly stated, according to a report by Carrier Management, that they do not trust Tesla’s self-driving technology or its safety claims, citing concerns about labelling consistency and internal pressure around performance metrics.

What the Tests Show

The numbers from independent testing are striking. AMCI Testing, an automotive testing firm, reported that in 1,000 miles of FSD testing, more than 75 human interventions were required to prevent unsafe behaviour — roughly one intervention every 13 miles. By comparison, US police-reported crash data suggests human drivers average about one crash every 165,000 miles. Critics note that gap is not small; it is orders of magnitude. That said, this is an independent test result, not an official government safety benchmark, and an intervention is not the same as a crash.

Tesla pushes back with its own figures. According to the company’s Safety Report, supervised FSD records one major collision about every 5.3 million miles — which Tesla frames as roughly seven to eight times fewer collisions than average human drivers in comparable conditions. Those numbers are self-reported and have not been independently audited to a standardised methodology.

Early data from Tesla’s unsupervised robo-taxi pilot, discussed by safety analysts drawing on US incident reporting rules, suggests around 800,000 miles driven with roughly 14 to 15 reportable incidents — one incident every 57,000 miles or so. Analysts describe that as about four times worse than average human drivers in similar urban conditions, though most events are reported as low-speed, minor collisions. These are early-phase figures, not a settled performance baseline.

The Human Factor

A peer-reviewed study of 103 Tesla Autopilot and FSD Beta users found something that goes beyond raw statistics. FSD Beta, the researchers concluded, is unfinished technology that requires constant supervision, increases driver stress and workload, and can be “inherently unsafe” due to unexpected system behaviour at critical moments.

Users adapted — and not always safely. Some placed weights on the steering wheel to fool the driver-monitoring system. Others used FSD on roads outside its intended design domain. Over time, some became complacent, their attention drifting in ways that left them poorly placed to intervene when the system did something wrong.

David Zipper, a visiting fellow at Harvard Kennedy School who has written extensively on autonomous vehicle policy, said: “The name ‘Full Self-Driving’ creates an expectation that the technology cannot currently meet, and that expectation is dangerous.”

Independent road testers have documented FSD driving on the wrong side of the road, making incorrect lane choices, and performing sudden, surprising manoeuvres — behaviours that successive software updates have not eliminated entirely, according to InsideEVs’ 2024 safety review and other road tests.

Where Regulators Stand

Regulators in the US and elsewhere have opened investigations into Tesla’s driver-assistance features following collisions, reported red-light violations, and lane errors. Software-only over-the-air updates have been used repeatedly to modify FSD behaviour after incidents and investigations — a pattern that reflects both the flexibility of the platform and the frequency of identified problems.

The branding debate is not going away. Safety advocates argue that “Full Self-Driving” is misleading for a Level 2 system, potentially encouraging drivers to treat it as something it legally and technically is not. Tesla maintains that in-car warnings and documentation make the driver’s responsibilities clear.

There is, as yet, no independent peer-reviewed consensus that FSD is statistically safer than human drivers across all conditions. Safety comparisons currently rely heavily on Tesla’s own data and selected independent tests with differing methodologies. The claim in this week’s tweet — that FSD can predict and respond to bad driving “often before it happens” — remains unverified as a quantified statistic. No independently validated metric currently confirms how frequently that anticipatory capability actually prevents harm.

What This Means for Kent Residents

Tesla FSD is not legally approved as an unsupervised self-driving system on UK roads; any Kent motorist using FSD or Autopilot on the M2, M20, or local A-roads remains fully responsible under UK road traffic law and must stay attentive at all times. Kent’s proximity to Dover and the Channel Tunnel means high traffic density and complex junctions — exactly the conditions where independent tests have documented FSD struggling — making over-reliance on the system a particular concern. Residents considering a Tesla with FSD, priced in the thousands of pounds as a software option, should be aware that safety claims currently rest largely on Tesla’s own self-reported data rather than independently audited figures.