Last week in a seminar we discussed a text that was largely about sexual violence, including mass rape during war. Heavy stuff.
One student admitted they had not read the text but worked off a ChatGPT summary.
They had no idea the text was about sexual violence. ChatGPT withheld that information.
This wasn’t just a minor error nor a typical LLM hallucination.
About a third of the text, arguably its most important part, went completely ignored because it didn’t match OpenAI’s content policies