There are other places where some process has an error rate and you make up for that error rate by doing the work more than once and then comparing results. For example, I've heard in a video that satellites and other space craft often have 3 or 4 processors and compare the results to make sure there were no errors due to radiation. Similarly, we have RAID arrays that store data multiple times because disks can fail. So, even if AI has a failure rate of like 20%, maybe you can make up for that by running the same prompt multiple times with slight variations or with different models, comparing the results and choosing the best.