TL;DR
A Reuters investigation found that 7 out of 9 Tesla data labelers would not work in FSD. They routinely watched on camera as the system sped up and crashed.
Reuters interviewed nine former Tesla data taggers and one former self-driving engineer about their views on Tesla’s fully self-driving mode. Seven of the nine data specialists said they would not ride in a Tesla that ran on FSD.. One said they would not travel in a Tesla robotaxi”if you paid me“
“We’ve all seen it fail,” a source told Reuters. The former autonomous vehicle engineer agreed: “Definitely don’t trust Elon on this.“They were referring to Musk’s statement that Tesla vehicles are ready for”safe without supervision” walks.
The data taggers’ job was to review hours of FSD images and train the vehicle’s software to avoid past mistakes. They had direct access to terabytes of proprietary driving data. At least five told Reuters they routinely saw clips of Teslas driving above the speed limit while operating FSD.
Engineers and managers treated the problem of speeding as a low priority. Edge case problems, such as unusual road configurations or unusual lighting conditions, received more attention. The priority of normal speeding, which affects all journeys and all roads, has been removed.
The 💜 of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The investigation comes as Tesla has expanded FSD availability to new markets. Tesla confirmed FSD availability in China last weekalthough it is not yet clear whether conventional consumers can still activate the system. The FSD (Supervised) system is classified as Level 2 and requires constant driver attention. A fully autonomous, unsupervised version is being tested solely on a fleet of robotaxis in Austin, Texas.
The last few months have produced a number of FSD-related incidents. Teslas operating with FSD have entered lakes, bridges and the path of oncoming trains. These are the incidents that reached media coverage. Testimony from data taggers suggests that the internal images contain a much larger catalog of failures.
The gap between Musk’s claims and the system’s performance has been a persistent problem. Musk has promised fully autonomous driving repeatedly since 2016. Each deadline has passed without delivery. The company’s robotaxi service in Austin operates in a geofenced area with security drivers available remotely.
Flood-related Waymo closures this month showed that even the most advanced autonomous driving systems have failure modes under routine conditions. Tesla’s approach differs fundamentally from Waymo’s: camera-only perception versus multi-sensor fusion, and a consumer vehicle repurposed for autonomy versus a purpose-built robotaxi.
The data labeler’s testimony is significant because these workers are the closest to the raw performance data. They don’t see marketing materials or earnings call projections. They watch hours of video showing exactly how the software behaves on public roads. Seven out of nine would not travel on the product they helped build.
Tesla did not respond to Reuters’ request for comment. The company has previously said that FSD (Supervised) requires active driver supervision and that its safety statistics show the system performs better than human drivers per mile. The former engineer interviewed by Reuters questioned those statistics.
The research raises a question that Tesla’s regulatory filings and marketing don’t address: If the people who train the AI don’t trust it, why should the people who ride in it?
