AI won’t “slowly” swallow white-collar work. It will do so all at once, once it crosses a certain critical threshold.
Not just because of raw reasoning ability, but because *once a model’s probability distribution can be made to converge reliably—so that it can repeat specific operations with consistency*—it will suddenly be able to take over a huge portion of routine white-collar tasks.
We don’t know exactly when that threshold will be reached.
But compared to GPT-4, which often produced silly hallucinations, GPT-5.1 already feels strikingly stable.
If model reliability continues to improve at this pace, it’s hard not to feel that the day AI “swallows” those tasks might not be far off.
I’m actually Japanese and can’t write English well, so I wrote the comment in Japanese and had GPT-5.1 translate it for me.
What I posted is still my own thinking — I just let the model handle the English part.
We don’t know exactly when that threshold will be reached. But compared to GPT-4, which often produced silly hallucinations, GPT-5.1 already feels strikingly stable.
If model reliability continues to improve at this pace, it’s hard not to feel that the day AI “swallows” those tasks might not be far off.