I would argue for an AI (I wouldn't call these language models that, but fine) to be controlled they need to understand (or have similar “reasoning” properties)

I would argue for an AI (I wouldn't call these language models that, but fine) to be controlled they need to understand (or have similar “reasoning” properties). Just because we anthropomorphize these models doesn't mean there is actually any reasoning skills behind them. You can't control these language models in the same sense that you can't control random number generators. You can only change the output once it's done.