See the biggest issue is that there isn't an easy way to test any hypothesis here.

If I had to speculate I could see both turning out to be true. There are probably some pedophiles whom AI CP will help handle the urge, and some for whom the readily available content will make actual abuse more morally acceptable. But then again, we'll probably never know for sure unless we find some criteria like in your nice battery example. Criteria such as "is the building on fire" give you quick and near-immediate feedback on whether or not you've been successful.

The discussion reminds me of the never-ending debate on whether drugs should be legal though. If there should be tests with AI CP, could there be a setup similar to that of supplying recovering heroine addicts (and only them) with methadone? This would allow the tests to be conducted in a controlled environment, with a control group and according to reproducible criteria.