Both are forms of censorship and both are bad, but personally I’d rather an AI shut up than slip in misinformation. If someone is going to take AI at its word (which people shouldn’t do, but you know how people are) then the misinformation will be taken as truth, but a message saying “I can’t discuss such things” will make someone turn elsewhere for answers.
It depends on what you consider worse. Deepseek will outright refuse to answer a question, but other models will happily answer with poisoned data.
See for example, grok translations of neutral questions about israel completely changing the questions: https://www.instagram.com/p/DXpK0BvFOUn/
and israel investing in influencing the output of chatgpt: https://www.ynetnews.com/tech-and-digital/article/rj00kxqzaxx
Both are forms of censorship and both are bad, but personally I’d rather an AI shut up than slip in misinformation. If someone is going to take AI at its word (which people shouldn’t do, but you know how people are) then the misinformation will be taken as truth, but a message saying “I can’t discuss such things” will make someone turn elsewhere for answers.