Pokémon Snap first came out today 20 years ago. It is an on-rails shooter game. You’re in a car that drives around an island filled and your goal is to take pretty pictures of the pokémon that live there. It was one of my favourite games when I was little. Its relaxing gameplay and cute jokes still managed to put a smile on my face when I played the game again recently for the first time in years.
It takes only a couple hours to take a picture of every pokemon in the game, though with high variance since some pokemon are really tricky to find. The game feels quite modern despite its age.
At the end of every course you can show a number of photos to Professor Oak and he will rate their quality on a couple different aspects, like how well the pokémon fits in the frame and if they are in a nice pose. The amazing thing is that algorithm almost exactly matches my own opinion of which photos are good. You don’t see that often in algorithms.
You have probably heard the latest AI news cycle, this time the news is about a language model. Let’s forget for a second about the minutiae of how dishonestly the achievement was presented and whether or not the software should be released to the public, other people have done that better than I ever could. I want to point at a cute problem related to authorship and suggest a possible solution.
In a future where most text will be governmental and commercial propaganda written by software, would people want to make sure they only read text written by other sentient beings? Or would at least old people like future me want that? And will that be a rational preference, or just anti-robot prejudice?
Let’s suppose that knowing whether a text is human-written is desireable. How could the future people go about doing that, and can we do anything now to help them?
The interesting case is in verifying that an anonymous self-hosted blog is human-written, or anonymous comments on such a blog. Most other cases seem to be trivially solvable either by having national governments verify digital identities or using interactive protocols (like Turing tests and captchas).
I’ll describe one possible avenue, useable only by anonymous internet users who act now, and possibly only for low stakes. The trick is to generate a pgp key and use it to sign every bit of content you create. The uniqueness of the signature (no two people can claim the content for their own) is established jointly by all internet archiving institutions, who also implicitly timestamp all data. All well-written text from before 2020 has been written by humans, so if you can claim authorship on a good share of that text then you are solid.