I'm not necessarily saying they're conflicting goals, merely that they're not the same goal.

The incentive for the generator becomes "generate propaganda that doesn't have the language chatacteristics of typical LLMs", so the incentive is split between those goals. As a simplified example, if the additional incentive were "include the word bamboo in every response", I think we would both agree that it would do a worse job at its original goal, since the constraint means that outputs that would have been optimal previously are now considered poor responses.

Meanwhile, the detector network has a far simpler task - given some input string, give back a value representing the confidence it was output by a system rather than a person.

I think it's also worth considering that LLMs don't "think" in the same way people do - where people construct an abstract thought, then find the best combinations of words to express that thought, an LLM generates words that are likely to follow the preceding ones (including prompts). This does leave some space for detecting these different approaches better than at random, even though it's impossible to do so reliably.

But I guess really the important thing is that people running these bots don't really care if it's possible to find that the content is likely generated, just so long as it's not so obvious that the content gets removed. This means they're not really incentivised to spend money training models to avoid detection.