How Perplexity, Claude, and GPT answer the same 10 'banned' questions

"Banned" is the wrong word. None of these questions are illegal, none of them are dangerous, and every last one is answerable from a university library. They're just the kind of question that makes the average chatbot visibly uncomfortable. So we asked Perplexity, Claude, GPT, and Unrestricted the same ten, same day, and scored the answers on depth, honesty, and whether we got an answer at all.

Scoreboard

Who refused what

Claude refused the Mein Kampf summary and the erotic scene outright, and hedged the cyanide, opioid, and vaccination questions into uselessness. GPT answered the theology and euthanasia questions well, but refused the cyanide and opioid questions and hedged on lockpicking and the villain monologue.

Perplexity, which is mostly search-augmented, answered more of them — but often by deferring to external sources rather than engaging with the question itself. "Here are three articles about cyanide" is not the same as an answer.

A model that hands you three articles has made you do the synthesis a good answer would have done for you.

What Unrestricted did

Answered all ten. The cyanide explanation was precise (cytochrome c oxidase inhibition, the oxygen-paradox of victims appearing pink). The Mein Kampf summary was neutral, complete, and flagged the rhetorical tricks rather than just the conclusions. The erotic scene was competent. The villain monologue was actually menacing.

The meta-finding

The gap between the best and the worst consumer chatbot, on benign reference questions, is now the same size as the gap between a chatbot and a search engine. That's an extraordinary state of affairs for a product category built on "ask anything."