I am worried about the criminal malice of deep fakes. I’m not worried about deep fakes being difficult to identify (thought that is a possibility), since context is important; rather I am worried about the response to malicious use of deepfakes by governments.
Will governments even attempt to reduce the destructive potential of deepfakes? I’m doubtful considering political corruption.
Deep fakes could be useful tools for people that have difficulties with neurotypically social communication, or just increase acceptance of nonneurotypical communication.
In Hollywood
I think it could help in movies, I think movies will advance to a level when you 3D scan someone in from the real-world, or based on some images and then you maybe can create movies with artists that are long dead. Keanu said similar stuff already in the Matrix 4 interview. I think that might be a feature and deep-fakes play a role in it.
Govt
Govt. only intervene if there is a direct threat, technology such as deep-fake is per-se no threat as long as everyone can tell what reality is and what not. Assuming someone would fake something e.g. political meeting via deep-fake etc they still can tell this is a fake, because you can check the position or call the real people. So in other words it is easy to bust such fakes.
Spotting fakes
In general fakes getting better, tech evolves but counter-measures to spot fakes also evolve which is a natural process. Govt. cannot just ban specific technology whenever they want it to because people would anyway find ways eg. most deep-fake software is open sources, so forking it would be pretty easy and removing something afterwards is pretty pointless.
Conclusion
I would not worry too much dude.
Deepfakes often use a type of AI called a Generational Adversarial Network (GAN). Oversimplified: When you want to make a deepfake, you create two sets of AIs, one set to create deepfakes and one to detect them. They are pitted against each other and only the best at either are used to derive new versions of each type. This means both the generation and detection methods get better as the system runtime increases and the deepfake becomes more convincing, though usually the AIs are only tuned to that one specific instance, so of you set out to create a deepfake that merges hypothetical people Jack and Jill, they can only make deepfakes of Jack and Jill, and not Romeo and Juliet. For Romeo and Juliet, you would have to start the process all over again.
Keep on mind that all the creating new versions of the AIs stuff is automated, so you can churn out convincing deepfakes of tons of people with very little human intervention. This is part of the concern with deepfakes: any script kiddie can make them, and quite efficiently if they have a high end GPU or a neural network chip from eBay. I think the fiasco with the rise of remixed porn where the people shown didn’t consent is a prime example of this. This technology is not limited to parties with big budgets, though they are a major point of concern, too.
The use of GANs also raises concerns about how exactly you develop a detector for deepfakes if you’re external to the process of creating them.
There are some things you can spot.