I expect they’ll be refining how they do the censorship as time goes by, but at the moment it appears to be trivially easy to bypass it in DeepSeek. People observed it beginning to write an answer, but then erasing it after writing the word Tiananmen. I was wondering if it is just a simple string blacklist, and I suggested (I don’t use DeepSeek myself) asking for deliberate typos in the prompt to potentially get around it. An acquaintance took the idea and refined it, asking to use characters that look like letters but aren’t. Got a perfectly good, uncensored answer.
I expect they’ll be refining how they do the censorship as time goes by, but at the moment it appears to be trivially easy to bypass it in DeepSeek. People observed it beginning to write an answer, but then erasing it after writing the word Tiananmen. I was wondering if it is just a simple string blacklist, and I suggested (I don’t use DeepSeek myself) asking for deliberate typos in the prompt to potentially get around it. An acquaintance took the idea and refined it, asking to use characters that look like letters but aren’t. Got a perfectly good, uncensored answer.