It depends on how you use the engine. Obviously, picking its best move or nearly best move when they're close in value, will get you caught.

But suppose you sample from the engine's ranked moves to have some decent ACPL (average centipawn loss) error rate. Just a slightly lower rate than you'd have on your own.

If you further bias your sample to moves to that look reasonable to you, then I don't see how you'd get caught.

Current engines may not have the right support for such sampling, but it wouldn't be hard to implement, e.g. with a private fork of Stockfish.

I can imagine training ML models that can imitate human play closely enough and even your play specifically to increase the odds you win.

Detecting a stockfish moves over enough sample size is easy sure, but detecting a engine which is designed to imitate human play not make the best move everytime is not easy with number of moves a human would play in their lifetime.