Try using an early net of LCZero. You used to be able to adjust how many "rollouts" it plays with, and what generation net to play against. Since it's a neural net, it has the "intuition" of a 1000 elo player, or 1500, etc. I.e. it does not play expert then blunder. Hopefully they get the play.lczero.org website back up and running sometime. I didn't have too much success getting the other links to work... https://blog.lczero.org/2018/11/where-to-play-leela-online.h...

Let me know if you find an easy way to play "human" level LCZero AIs.

Edit: Main project website: https://lczero.org/

I see a practical use of this when you are offline and without a chessboard or someone else.

Otherwise, you can play online quite easily. For a game I only wait for a few seconds on lichess.org

>randomly making a deliberately terrible move every now and then.

To be fair though, that is kind of how I end up playing chess. I'll do pretty well and pay attention, then I end up losing track, making a stupid mistake and doing something that should have been obviously bad and losing.

I had Chessmaster 8000 and the engine behind it was always really good at mimicing realistic playstyles. It had presets, and you could control things like pawn weakness, willingness to trade, king safety, and preferred pieces.

Modern chess engines usually have a notion of "contempt". I've found that if you turn contempt way up and limit the number of concurrent lines, in combination with adjusting the target rating, you can get a reasonably good approximation of real play. At casual (< 2100 maybe?) levels of play, this simulates the human tendency to make blunders while fixating on a plan.