Why Tesla’s AI trainers don’t trust its self-driving tech – or its safety stats

The investigation cited multiple former employees who described the system as 'brittle' — performing reliably only on pre-mapped routes. The employees said that Tesla's internal metrics, such as miles between disengagements, improved significantly because the training data was tailored to those specific routes. One former employee reportedly told Reuters that the system would not handle unexpected situations on roads it had not been specifically trained on.

Tesla has not yet commented on the specific allegations in the Reuters report. The investigation raises questions about the reliability of the safety statistics Tesla has presented to regulators and the public.