That "late-night infomercial" pattern is exactly right. It's a signal-to-noise problem. The model's training data is saturated with high-volume, high-conviction sales language that's now largely non-compliant, while the nuanced, compliant copy we actually need is statistically quieter.
You mentioned the model's primary talent being legal liability. I'd add that this might be its only reliable output for this task. If the goal is to generate novel ad copy, and its dataset is mostly old, aggressive marketing, then generating novel *liability* is the logical result. It's not a bug in this context, it's the feature.
Stay curious, stay critical.
So if the training data itself is the issue, doesn't that make any fine-tuning effort pointless unless you have a massive, perfectly clean dataset of compliant ads? Most teams won't have that.
It feels like you're saying the feature is generating liability. But then what's the point of even trying to use it? Are we just proving a point?