Hear me out: LLMs should store their responses so that when someone asks them "did you write this?", even with modifications of the response they can confirm or deny. Since they are costly to operate, it will cut down on their abuse.
There is the halting problem and the decidability problems. You can approximate certain things using a perceptron/neuron floating point configuration, but they are only approximates; and do nothing towards addressing either of those problems.
Kidding aside, no one knows what exactly we do differently that allows for complex thought, context switching, or conflict resolution that works around those aforementioned problems.