Cryptography isn’t really the hard part here. The hard part is that for a computer to be useful it needs a way for humans to use it, and you can’t really allow humans to use a computer while preventing an AI from using it. For example, if you allow humans to use keyboards, an AI could use a trivial robot to press keys on the keyboard; and if the computer has a visible output, an AI could read it with a camera.
And that’s assuming you completely control the supply chain of all computer parts and manage to make them tamper-proof (also extremely difficult) so the AI can’t do something easier like send USB inputs.
The strongest case of this working that I’m aware of is video DRM, where a video player can display a movie only cryptographically verified and somewhat tamper-proof displays, but it would be trivial to let an AI watch a DRM-protected movie by pointing your webcam at the screen (and someone more determined could take the monitor apart and read the signal after decryption).
Cryptography isn’t really the hard part here. The hard part is that for a computer to be useful it needs a way for humans to use it, and you can’t really allow humans to use a computer while preventing an AI from using it. For example, if you allow humans to use keyboards, an AI could use a trivial robot to press keys on the keyboard; and if the computer has a visible output, an AI could read it with a camera.
And that’s assuming you completely control the supply chain of all computer parts and manage to make them tamper-proof (also extremely difficult) so the AI can’t do something easier like send USB inputs.
The strongest case of this working that I’m aware of is video DRM, where a video player can display a movie only cryptographically verified and somewhat tamper-proof displays, but it would be trivial to let an AI watch a DRM-protected movie by pointing your webcam at the screen (and someone more determined could take the monitor apart and read the signal after decryption).