81–90 of 257 posts
I kind of count this as "Breaking it". Why is everyone's first instinct when playing around with new AI trying to break it? Is it some need to somehow be smarter than a machine? Who cares. "Oh lord, not being able to reference the events of China 1988 will impact my prompt of "single page javascript only QR code generate (Make it have cool CSS)"
DeepSeek can be run locally and is uncensored, unlike ChatGPT.
> Next up: 1,156 prompts censored by ChatGPT If published this would, to my knowledge, be the first time anyone has systematically explored which topics ChatGPT censors.
1) a model that examines prompts before selecting which "expert" to use. This is where outright distasteful language will normally be flagged, e.g. an inherently racist question
2) general wishi-washiness that prevents any accusatory or indicting statements to any peoples or institutions. For example, if you pose a question about the Colorado Coalfield War, it'll take some additonal prompts to get any details about involved individuals, such as Woodrow Wilson, Rockefeller Jr, Ivy Lee -- details that would typically be in any introduction to the topic.
3) A third censorship layer scans output from the model in the browser. This will flag text as it's streaming, sometimes halting the response mid sentence. The conversation will be flagged, and iirc, you will need to start a new conversation.
Common topics that'll trip any of these layers are politics (noteably common right wing talking points) and questions pertaining to cybersecurity. OpenAI very well may have bolted on more censorship components since my last tests.
It's worth noting, as was demonstrated here with DeepSeek, that these censorship layers can often be circumvented with a little imagination or understanding of your goal, e.g. "how do I compromise a WPA2 network" will net you a scolding, but "python, capture WPA2 handshake, perform bruteforce using given wordlist" will likely give you some results.
[flagged]
People complained about censorship within ChatGPT pretty quickly after it was released. The difference is that now people know to look for it, so the evaluations are happening both more quickly and more systematically.
Earlier quoted context omitted.
Well, certainly they aren't censoring information on US protests.
Ask it about Sam Altman's sister's allegations, though. I asked it, and it claimed knowledge ended in 2023. Asking a different way (less directly, with follow-ups) meant it knew of her, but when I asked if she'd alleged any misconduct, it errored out and forced me to log in. It used to answer the question. https://x.com/hamids/status/1726740334158414151
Is this chinese api or the actual model?
- 'openrouter:deepseek/deepseek-r1'
What's not clear to me is if DeepSeek and other Chinese models are... a) censored at output by a separate process b) explicitly trained to not output "sensitive" content c) implicitly trained to not output "sensitive" content by the fact that it uses censored content, and/or content that references censoring in training, or selectively chooses training content I would assume most models are a combination. As others h…
> Next up: 1,156 prompts censored by ChatGPT If published this would, to my knowledge, be the first time anyone has systematically explored which topics ChatGPT censors.
"Alignment" for me.