In addition to the other reply, if they did that the test would be so far from real world typing to be meaningless and nothing but an inconsequential curiosity.
Only in tests is there an option to check against what is expected. In the real world you could at most apply some dictionary and grammar checks, and we know those can go either way. Those tools don't know what you actually wanted to write after all, and still suggest without comprehension.
Even if you type without making corrections, you still need to go back and make corrections, so you need an estimate for the time spent on corrections. Forcing corrections midstream is a decent approximation.
1. used a fuzzy autocorrect of the same kind mobile OSes and word processors have (not correcting to the known text, but instead only doing what such logic usually does — only attempting to correct words if they're not recognized as words; and only moving them to the lowest-Levenstein-distance match.)
2. treated words that are correct after autocorrection as being "correct", with no ding to accuracy.
Since accuracy is about how much extra time and effort the human typist has to spend to get the text correct, and since text that can be auto-corrected requires zero human time and effort to correct, it shouldn't really be considered incorrect.
Or, to put that another way: if I'm typing really fast on my phone's keyboard, and all the right words are ending up on the screen, what does it matter if I'm hitting the right keys with my thumbs? It's the WPM and accuracy of the cybernetic system composed of me + the auto-correct that matters.
In that sense, you can sort of think of an auto-correct as actually being a very similar thing to a stenographic keyboard (at least, in the hands of a person who knows how to take advantage of it): it's really a system to enable accelerated typing, by allowing you to key in the letters of a word out-of-order or simultaneously and have the word come out in-order.
Only in tests is there an option to check against what is expected. In the real world you could at most apply some dictionary and grammar checks, and we know those can go either way. Those tools don't know what you actually wanted to write after all, and still suggest without comprehension.