Deepfakes are for deceiving humans. There are still way too many artifacts in the outputs that are easy to detect in code.
The problem comes that even with extremely good deepfake detection, political operations can still use the deepfakes to fool the average voter into believing that is it real, irrespective of any successful analysis. You and I might not be fooled but we don't matter if 10000X people are fooled and affect our lives anyway in a negative way.
For propaganda purposes, the deepfake needs to be neither real-time, deep, nor fake, nor does it need to exist at all. A false message with zero supporting evidence repeated endlessly by high-profile influencers is a million times more effective than a single bombshell video ignored by mainstream media.
You don't need to detect a deepfake in the same time as it takes to generate one, do you?
But you may have to go through orders of magnitude more material
If you get a Zoom call from your CFO telling you to issue a money transfer to that bank account, that it is 100% legit, how much time do you have to detect a deepfake?
Does your company instantly execute payment instructions to new supplier accounts based on a remote verbal instruction?