This isn't true. The principle in general is that you want to detect situations of danger, minimizing false positives and false negatives, with your pay off function probably valuing false negatives as much worse than false positives. With more sensors, the worst case scenario is that it doesn't improve the payoff function and you simply don't use the sensor information. But in the best case scenario the extra sensor can be used to massively increase your payoff - your camera gets confused by a shadow for example, but your radar doesn't detect shadows, so you can eliminate those false positives. Since the general principal to solve self-driving has been to train AI to process sensor input to make decisions, the worst case scenario is you end up with a neural net that literally just doesn't look at the other sensor data (assuming you're training your AI reasonably, if you aren't then maybe you shouldn't be working on an incredibly difficult problem with safety critical applications)
Tesla has essentially bet that they can get their optical systems so good that other sensors like Lidar do hit that situation where they offer nothing in terms of improving the pay off, but every piece of evidence we've seen so far has indicated that they haven't achieved that, and this is many years after Tesla had been confident that they would acheive it.