Gannett was publishing AI-generated sports stories in its local markets, and they were awful. Go read the story, it gives you plenty of threads to pull at to find examples of failure.
Companies like Narrative Science and tools like Heliograf were the historical harbingers of this trend, but all of their auto-generated content was based on templates.
For example, we could write a sentence about who won by comparing the scores of the two teams:
[if team1_score > team2_score]
[team1_name] beat [team2_name] [team1_score] to [team2_score].
[else]
[team2_name] beat [team1_name] [team2_score] to [team1_score].
That's 100% fake code, but the idea is that everything was about filling in the blanks, not about generating novel content. With the modern batch of generative AI tools, you're turning the AI loose on the data and letting it figure out what's important and what's not, along with exactly how to phrase everything.
With the unpredictable nature of these tools, it's not a terribly good idea, and I'm certain we'll continue to see more screw-ups of this sort.
We perfected the technology years ago!! Why can't we just have templates back!!