Comment by beloch

What would LLM's make of normal human conversations if they had access to everything you say (Just wait!)? Think back to the last time you hung out with a group of friends, either in person or online. How much of what was said would an AI bin into positive, negative, or neutral categories?

Rather a lot of what is said in any given social circle has to do with complaining. It's very common for people to point out something that is viewed as bad by everyone. Then the group commiserates and bonds over that. Even though an AI might consider such complaints negative, there might be a positive effect on people feeling heard and supported by a like-minded group.

For this reason, I'd take OP's results with a modicum of salt. Human interaction doesn't have to be all rainbows and unicorns to have a positive psychological impact. As with in-person interactions, I suspect a significant portion of what OP's LLM's described as negative might just be humans bonding through complaint among peers.