So far I tested deepseek r1:7b and llama3.2 post on the fediverse and on detecting AI content (I have no shortage of AI content). Seems like llama3.2 takes preforms worse at detecting AI content compared to deepseek but works quicker. I’m not working with any graphics cards. Deepseek model be really good at generating replies with the right prompt, but it does take several minutes to run. No one has said anything about AI text I submitted other than reply I made to the wrong person and people are even sharing and liking the AI produced text
Doing this research transparently is a good idea, it will only be a matter of time before other actors will try and use LLMs to influence lemmy. Like spam, bots and farms, this is a reality the fediverse will have to deal with eventually. Better start preparing now.
It’s interesting that the deepseek AI can blend in. It made a post like it really visited a waterfall. It didn’t say anything too specific as to make it obvious it wasn’t there(maybe that the weather was good in this time of year in March). So it will require looking at multiple posts or maybe even interacting to spot the bot.
These boys can definitely blend in and then run script that can push an ideology. I won’t make more than 1 since I don’t have the resources