The Cost of Arriving Early
The best sentence in a first draft is almost never the one you sat down to write. It arrives sideways, mid-paragraph, while you're explaining something else, and a thought you didn't plan for muscles its way onto the page. You notice it because it doesn't match the register around it. Everything else is doing its job. That one line is doing something else.
That sentence needed room to happen. Not time in the calendar sense, but the particular kind of room where the sentence you meant to write has finished and the next one hasn't started, and nothing is standing in between telling you what goes there. The digression gets born in that opening. If something fills it before your own thought arrives, the thought doesn't get cancelled. It just never happens. You don't notice the absence, because there's nothing there to compare it to. The sentence that would have existed doesn't leave a hole shaped like itself.
This is the actual damage predictive text does, and it has nothing to do with accuracy. A tool that finishes your sentence wrong is easy to catch and delete. A tool that finishes your sentence plausibly is the one that costs you something, because plausible is exactly the threshold your own attention uses to decide whether to keep looking. You're not weighing the suggestion against the sentence you would have written. You're weighing it against nothing, because the sentence you would have written hasn't happened yet. It needs another two or three seconds of not-having-an-answer before it shows up, and the completion arrives in the first half second, closing the opening before the actual thought has cleared its throat.
Call it the arrival problem. The machine's guess isn't usually bad. It's good, in the narrow sense of being grammatical, on topic, statistically the kind of thing that sentence tends to end with, and that's exactly the problem. A bad guess you'd reject on sight. A good-enough guess you accept on reflex, and reflex is faster than the process that produces a real digression. Real digressions are slow. They need the sentence to stall, the writer to sit in that stall for a beat, and some unrelated thought, a memory, a stray association, an argument with yourself, to wander across before the next sentence closes the loop on its own. Autocomplete's whole value proposition is closing loops faster, which puts it in direct competition with the one mechanism that produces anything worth reading.
This isn't only a writing problem. The same shape shows up in conversation, in code, in the reply you almost sent before the suggested reply loaded first. Anywhere a person produces language under time pressure, an assist that arrives early is doing more than saving keystrokes. It is pre-empting the search before the search starts. A search that gets pre-empted doesn't fail loudly. It stops running, quietly, and you keep typing, and the draft reads fine, and nobody, including you, can point to the sentence that isn't in it.
None of this argues for turning the tools off, and it isn't an argument against speed generally. It's an argument for noticing which parts of the process actually need the delay. The pause between one sentence and the next isn't dead time to be optimized away. Some of it is where the only original thing in the piece is quietly forming, and it needs you, and whatever's helping you, to be a beat slower than either of you is capable of being.
The tools are rarely wrong enough to worry about. They're early, constantly, and early is the failure mode nobody's built a warning light for.