My understanding is that they don't bring the service to exactly the SLO... To prevent overreliance on a service, it can be sufficient to introduce some level of failure, which may still be well above the SLO.
One major issue and you stop artificially inserting errors, which are inserted at a rate such that you could turn off the error inserted within some timely manner and still stay within budget.
Why ride the line? It'd take one major issue then you're way over you error budget?
For testing/simulations I can see why you'd introduce the errors.