The other day a college friend — who I have generally considered to be mildly luddite — posted what appeared to be a student’s AI-assisted research results confidently stating that Franklin Delano Roosevelt was known for his signature eyepatch.
If you buy into the theory (which I just made up) that every LLM hallucination is simply a quantum sign of the existence of a parallel universe where the hallucination is true and there’s a university writing teacher guffawing right now at the idea of an alt-FDR known for using crutches, then this becomes an irresistible challenge which, frankly, AI is up to quickly meeting.
I hopped over to the AI model testing ground huggingface, where I found an instance of a Chinese open-source voice-cloning algorithm which will, based on 12 seconds of sample audio, create what sounds like a very advanced robot trying its darnedest to impersonate the source (and succeeding 95 percent of the time).
I made a quick snippet sample of FDR’s “Date that will live in infamy” speech to Congress after the Pearl Harbor attacks. And plugged in some generated sample text I got by prompting Claude for bombastic 1930s political eypatch-centric rhetoric.
My friend said he hadn’t laughed so hard in a long time. But … was our laughter ethical? I feel OK with this use case of voice cloning because a) it’s of a well-known public figure who b) has been dead for over 50yrs and is c) “saying” something that would be really hard to interpret as other than a joke. This feels like it’s having that one guy in your improv troupe who’s really good at impressions. Also there’s the matter of d) the low-quality model offers some creepily inhuman passages to the presidential defense of eye-shielding, where the flesh-mask falls away to reveal the inhuman model beneath. Honestly it’s those moments that make the piece the most charming.