Why AI posts drag on
RLHF trains models on human preferences, and humans conflate verbosity with quality, a measured length bias. Audiences decide within one to three seconds whether to keep reading, so the training pushes output in exactly the wrong direction.
29 June 2026 · Note · Applied AI
Longer responses, preferred by human raters
1 to 3 seconds to decide whether to continue
Humans conflate verbosity with quality, so models learn to belabor the point. The audience decided seconds ago.
Why do so many AI generated posts seem to drag on and on?
One method used for training AI is called RLHF, or reinforcement learning from human feedback. Quite simply, this is when humans choose the AI response they prefer, which is then used to guide future responses from the model.
A study from Seoul National University found that this training method creates “length bias”, as humans tend to “conflat[e] verbosity with quality”. By preferring longer responses, we train models to produce more of them.
This directly conflicts with trends in content consumption, where people decide within 1-3 seconds whether to continue with a piece of content. Belaboring your point actively reduces the chance that your audience hears it.
Not all AI content is slop, but rambling content that uses lots of words to say nothing is.
