This was comments, not posts. They were using a model to approximate the demographics of a poster, then using an LLM to generate a response counter to the posted view tailored to the demographics of the poster.
You’re right about this study. But, this research group isn’t the only one using LLMs to generate content on social media.
There are 100% posts that are bot created. Do you ever notice how, on places like Am I Overreacting or Am I the Asshole that a lot of the posts just so happen to hit all of the hot button issues all at once? Nobody’s life is that cliche, but it makes excellent engagement bait and the comment chain provides a huge amount of training data as the users argue over the various topics.
I use a local LLM, that I’ve fine tuned, to generate replies to people, who are obviously arguing in bad faith, in order to string them along and waste their time. It’s setup to lead the conversation, via red herrings and other various fallacies to the topic of good faith arguments and how people should behave in online spaces. It does this while picking out pieces of the conversation (and from the user’s profile) in order to chastise the person for their bad behavior. It would be trivial to change the prompt chains to push a political opinion rather than to just waste a person/bot’s time.
This is being done as a side project, on under $2,000 worth of consumer hardware, by a barely competent progammer with no training in Psychology or propaganda. It’s terrifying to think of what you can do with a lot of resources and experts working full-time.
This was comments, not posts. They were using a model to approximate the demographics of a poster, then using an LLM to generate a response counter to the posted view tailored to the demographics of the poster.
You’re right about this study. But, this research group isn’t the only one using LLMs to generate content on social media.
There are 100% posts that are bot created. Do you ever notice how, on places like Am I Overreacting or Am I the Asshole that a lot of the posts just so happen to hit all of the hot button issues all at once? Nobody’s life is that cliche, but it makes excellent engagement bait and the comment chain provides a huge amount of training data as the users argue over the various topics.
I use a local LLM, that I’ve fine tuned, to generate replies to people, who are obviously arguing in bad faith, in order to string them along and waste their time. It’s setup to lead the conversation, via red herrings and other various fallacies to the topic of good faith arguments and how people should behave in online spaces. It does this while picking out pieces of the conversation (and from the user’s profile) in order to chastise the person for their bad behavior. It would be trivial to change the prompt chains to push a political opinion rather than to just waste a person/bot’s time.
This is being done as a side project, on under $2,000 worth of consumer hardware, by a barely competent progammer with no training in Psychology or propaganda. It’s terrifying to think of what you can do with a lot of resources and experts working full-time.