With a sufficiently effective language model it would be fairly easy to bridge this model, by letting the text direct humans on the other side.

    Hypothesis: <AI writes this>
    Results: <human observations>
    <repeat>