TL;DR: Tesla self-driving tech is becoming less safe per mile, according to Tesla’s own data. Q1 2025 was 2.5% worse than Q1 2024. Q2 2025 was 2.8% worse than Q2 2024.
There’s some rough news released alongside Tesla’s Q2 2025 financial results, maybe even worse than the plummeting profitability and the total lack of guidance as to when the good days could return.
Year-over-year for the last two quarters, Tesla has been reporting more crashes per mile for their self-driving technology.
That’s strange. We expect fast-moving technologies like self-driving to get better, not worse. So what gives?
Turns out, the declining safety of Tesla’s self-driving tech follows a risky design decision.
Major Updates, Major Risks
At the end of 2024, Tesla completed a radical refactoring of their self-driving technology.
According to Tesla’s own release notes for the first phase of the change:
“FSD Beta v12 upgrades the city-streets driving stack to a single end-to-end neural network trained on millions of video clips, replacing over 300k lines of explicit C++ code.”
By year’s end, the same approach would be applied to highway driving. The same design philosophy is also applied to AutoPilot, Tesla’s free ADAS system.
This might sound exciting if you’re a science-fiction writer or a venture capital bro who is looking for a stock to pump. But actual artificial intelligence experts? Probably pulling their hair out. Here’s why.
The Devil Is in the (Technical) Details
AI experts know that neural networks, while capable of interpreting an incredible variety of scenarios from training data, they’re fundamentally hard to predict. Replacing 300,000 hard-coded lines of programming with an AI model is a fundamentally risky proposal. Those ~20MB of data can stand between drivers and a fiery end if their car’s AI mistakes an overturned truck for an open highway or makes an otherwise inexplicable mistake.
That’s because, in certain applications, you really want hard-coded behavior. For instance, when driving, it’s important to follow set rules, rather than replicate behavior you see in training data and videos. Hell, I might be in those training videos, and I sure don’t want any self-driving vehicle to drive like me.
Why even refactor the system? Well, in case you haven’t noticed, AI is the thing right now for increasing your stock price. Also, Tesla has designs on global domination, and manually hard-coding safe driving software for every traffic pattern on earth is a huge undertaking. There are national, regional, and even local variations, and each one is critically important to get right.
Far cheaper (and more likely to send the stock price to Mars) to rely on a flashy if fundamentally risky catch-all solution.
How Bad Is the Most Recent Safety Data?
There’s clearly some seasonality to Tesla’s self-driving safety data, which makes sense. But compare year-over-year, and the direction is unmistakable and coincides with the transition to neural networks that was completed at the end of 2024.
Q1 2025 was 2.5% worse than Q1 2024. Q2 2025 was 2.8% worse than Q2 2024.
Now, those drops aren’t devastating. The problem is that there are drops at all.
Take a look, here’s the data:
| Miles between crashes (higher is safer) | Q1 2024 | Q1 2025 | Q1 YoY Change | Q2 2024 | Q2 2025 | Q2 YoY Change |
|---|---|---|---|---|---|---|
| Autopilot in use | 7.63 M | 7.44 M | −2.5% | 6.88 M | 6.69 M | −2.8% |
| Autopilot not in use | 0.955 M | 1.51 M | +58.1% | 1.45 M | 0.963 M | −33.6% |
(Thanks to Electrek for compiling the data numerically from Tesla’s quarterly safety report so I didn’t have to!)
The problem is less regression, and more stagnation. As Waymo grows their service area while maintaining a state-of-the-art safety record, Tesla is moving in the wrong direction.
Tesla’s Stats Have Always Been Cooked
It’s worth mentioning: Tesla stats have always been dodgy as hell, and so this is Tesla admitting they have a problem, which means that the problem is likely worse than it appears.
After all, Tesla under-reports accidents by using a non-standard definition of what an accident is (IE, only airbag deployments or when other destructive safety measures deploy). A Tesla accident is fairly serious, while the NHTSA or Waymo have a much lower threshold to count a crash as a crash.
More on that here if you geek out over automotive safety stuff like I do: https://electrek.co/2025/06/16/bloomberg-most-embarassing-report-tesla-waymo-self-driving/
“Let’s Make the Most Misleading Benchmark Possible” – Tesla, Probably
Tesla also reports their self-driving safety data alongside a supposed “benchmark” of human drivers, but their methodology is comically flawed by including every single car on the road as their competition.
Here’s why that’s an obviously stupid way to view the data:
- EVs generally and Teslas specifically are built more recently than the average car. New cars are safer than old cars. Duh, right? EVs have grown about 10x in market share since 2017, and so EVs are far less likely to be busted old hoopties (aside from every single Cybertruck).
- Teslas are luxury vehicles. Expensive cars are driven by rich people. Duh. Rich people get into fewer car accidents.
- Drivers use driving assistance technology disproportionately on highways, with long stretches of simple driving. Drivers get into crashes more frequently within 5 miles of home, on local roads with their complex traffic patterns and intersections.
So, if you wanted to paint as misleading a picture as possible, you would do what Tesla has done, by presenting your ADAS safety statistics as compared to every car on the road while ignoring the significant differences between luxury EVs and every other car on the road.
That’s why we’re not parroting Tesla’s claim that their human supervised self-driving is something like 8x safer than the average driver. It just isn’t true in any meaningful sense.
Also, hah, if it was true, that means they’ve still got a ways to go before catching up to Waymo, which is actually 12x safer than a human driver (according to research from a leading European insurance firm).
The Long & Short Of It
Tesla’s self-driving safety is getting worse. Their own data says so. That data’s likely biased in their favor and it still shows regression. Meanwhile, competitors like Waymo are expanding while keeping crash rates ultra-low.
The likely culprit? Tesla’s risky move to go an “end-to-end neural network,” removing the hand-coded safety rails in favor of a black box AI. It may help with scalability. It may impress investors. But it’s a gamble that seems to be making the road less safe.
Stay tuned, we’ll be watching closely. But not too closely, probably going to give a couple extra car-lengths of following distance. You know, just in case the neural network gets any ideas.
