The power that GPT-3 gives to spammers has me worried as well. Anyone with sufficiently many IP addresses can now single-handedly kill any forum by spamming it

The power that GPT-3 gives to spammers has me worried as well. Anyone with sufficiently many IP addresses can now single-handedly kill any forum by spamming it with text which is at first sight indistinguishable from the average poster. Automatically generated spam could also be used to suppress discussion of political opinions or certain topics by drowning them in a sea of garbage comments, which seems highly problematic. We should come up with a solution to combat this. Here are a few ideas: - To a certain extend, the comment voting mechanism can be used to filter out comments which lead nowhere. If that stops working because comments are too good, that is not a problem. In the words of Randall Munroe: https://xkcd.com/810/ https://xkcd.com/810/ However, voting only works if the number of people voting on a certain comment outnumber the spam comments, so this will fail with too many spam comments. - Another solution would be to verify that every comment author is an actual human being. The GPG web-of-trust could be used for that, but this is untraceable for the average user. There also should be an additional layer of indirection between actual user identities and online identities to preserve privacy, but I am not sure if it is possible to have both privacy and limit a forum user to a single account so they can not circumvent a ban by simply creating new accounts. A trusted third party could solve this, but a distributed solution would of course be preferable. Maybe there is a smart cryptographic solution to this problem? - For the near future, every provider of GPT-3-based services should also provide a service to check whether some given comment has been generated by their model. This can be realized by hashing substrings of all generated output and storing the hashes in a bloom filter. This is not a long term solution since technological advancement will soonishly enable regular people to train similar models. - Training other neural networks to detect automatically generated text is not a solution because the generating networks can simply be trained to not be detected by the detecting networks. This is just a cat and mouse game. Does anyone have a better solution?