The problem I see here is logistical. AI is often a blank slate, dependent on the data that's fed into it to be useful. It's that data, and the way the model is tuned, that is important. We do control exports on certain grade encryptions, but how do we control export over tuning a machine model, or over input data for a ML model?