Tesla’s own AI trainers don’t trust ‘Full Self-Driving’ or its safety stats, Reuters finds

A Reuters investigation based on interviews with nine former Tesla data labelers, a former self-driving engineer, and 11 traffic-safety researchers found that Tesla's "10x safer than human" FSD safety claim relied on comparing Tesla airbag-deployment crashes to federal tow-away crash data, a mismatch a University of Michigan researcher recalculated to roughly 3x when normalized. The report also documents that Tesla extensively pre-mapped its Austin robotaxi zone and the Warner Bros. Cybercab demo route, doubling its Utah labeling staff to about 300 ahead of the June 2025 Austin launch, which still operates around 20 unsupervised vehicles.

Why it matters

The headline finding for operators isn't the safety stat. It's the mapping. Tesla has built its entire AV thesis, and a meaningful portion of its valuation, on the claim that it doesn't need the geofenced HD-map approach Waymo uses. Reuters now has sourced internal accounts of staff driving the Austin zone every night for weeks, labelers annotating curbs and stop lights by hand, and the Utah team doubling to ~300 specifically to prep the launch. That's the Waymo playbook with worse tooling, not a generalized self-driving stack.

The scaling math follows directly. Nearly a year in, Austin is still ~20 unsupervised cars in a hand-annotated zone, some with front-seat monitors. If each new metro requires hundreds of labeler-hours per square mile of operating domain, Tesla's cost-to-expand curve looks a lot more like Waymo's than like a software update push. The 3x vs 10x safety restatement matters for the NHTSA cases and the $243M Florida verdict; the mapping disclosure matters for anyone modeling robotaxi unit economics or competing capex plans.

Background reading

Related signals

← Read other signals from this week

Friday issue, by email. Free. No spam.

Background reading

Related signals

Subscribe