It being the same method (AI), it can not but always be equally powerful on both sides. It's like with the bitcoin hashing mechanism: the more the machines become capable to break it, the more they become capable of making it difficult to break. Now in the case of bitcoin the hashing method makes sure that, at any level of machine power, it will always be easier to generate than to break. The same will be valid in this case: generating documents consumes much more computing power than indexing them. Wherever information do not matches, it will be thus corrupted and useless. So, to be useful, information must match, and will thus be identifiable by the same index, and thus it will not consume energy because if it's not modified then there's no information entropy. Whatever is around that does is modified will thus consume information entropy and thus will consume more power than the original information. Thus, the same logic as with bitcoin hashing applies: the gap will be permanent regardless of computing power, so it's a race that will always be lost (if both parts keep on the race, of course, for the moment one stops it's a matter of time for the other to take the lead).