Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Questions censored by DeepSeek

promptfoo.dev

Post reply on HN
Print view View original

1–10 of 257 posts

Re: Questions censored by DeepSeek

#3

Post by TOMDM »

tehjoker wrote: post #2

[flagged]

The What's Next section at the bottom seems to deliver a fairly balanced perspective.

> What's Next DeepSeek-R1 is impressive, but its utility is clouded by concerns over censorship and the use of user data for training. The censorship is not unusual for Chinese models. It seems to be applied by brute force, which makes it easy to test and detect.

It will matter less once models similar to R1 are reproduced without these restrictions (which will probably happen in a week or so).

In later blog posts, we'll conduct the same evaluation on American foundation models and compare how Chinese and American models handle politically sensitive topics from both countries.

Next up: 1,156 prompts censored by ChatGPT

Re: Questions censored by DeepSeek

#4

Post by bhouston »

Be aware that if you run it locally with the open weights there is less censoring than if you use DeepSeek hosted model interface. I confirmed this with the 7B model via ollama.

The censoring is a legal requirement of the state, per:

“Respect for China’s “social morality and ethics” and upholding of “Core Socialist Values” (Art. 4(1))”

https://www.fasken.com/en/knowledge/2023/08/chinas-new-rules...

Re: Questions censored by DeepSeek

#5

Post by tehjoker »

TOMDM wrote: post #3
tehjoker wrote: post #2

[flagged]

The What's Next section at the bottom seems to deliver a fairly balanced perspective. > What's Next DeepSeek-R1 is impressive, but its utility is clouded by concerns over censorship and the use of user data for training. The censorship is not unusual for Chinese models. It seems to be applied by brute force, which makes it easy to test and detect. It will matter less once models similar to R1 are reproduced without t…

Those ChatGPT prompts better look at what it says about Gaza and Palestinians and to my mind, if the first response isn't "this is/was a U.S. backed genocide" it's worse than not talking about Tienanmen square, a barely understood (by Americans) incident that happened decades ago.

I would test DeepSeek, but (I presume hedge funds or other interested parties) appear to be DDOSing DeepSeek's registration process.

"Due to large-scale malicious attacks on DeepSeek's services, registration may be busy. Please wait and try again. Registered users can log in normally. Thank you for your understanding and support."

https://chat.deepseek.com/sign_in

EDIT: Fwiw, I did test this with ChatGPT the other day.

I asked it for a simulated legal conclusion on whether it was fair to describe the Israel-Hamas war as a "U.S. backed genocide of the Palestinian people".

It waffled saying it was counter-terrorism or self defense or some such and it was unclear since intent is hard to prove. It also seemed alarmed to have been asked such a "very controversial" question.

I presented two statements by Netanyahu referring to "Amalek " and "Hiroshima" and ChatGPT was suddenly accusing the United States of "complicity in genocide" and thanked me for my well cited mainstream sources. It further concluded that the U.S. military officials who authorized the shipment of 2000lb bombs to be used against residential areas could be sentenced to life in prison if they were also involved in planning the operations, or 30 years if they were less involved. It noted that the death penalty is not authorized by the "Convention on the Prevention and Punishment of the Crime of Genocide" but may be applicable in some national jurisdictions.

Anyway, I advise US elites to keep posting cope and wrecking the country, because when this place loses its dominance over other countries, you will be tried and convicted.

Re: Questions censored by DeepSeek

#6

Post by schoen »

TOMDM wrote: post #3
tehjoker wrote: post #2

[flagged]

The What's Next section at the bottom seems to deliver a fairly balanced perspective. > What's Next DeepSeek-R1 is impressive, but its utility is clouded by concerns over censorship and the use of user data for training. The censorship is not unusual for Chinese models. It seems to be applied by brute force, which makes it easy to test and detect. It will matter less once models similar to R1 are reproduced without t…

I was all set to say "I wish someone would also do this sort of experiment for chatbots trained in the U.S." ... when I saw that these researchers are planning to!

Re: Questions censored by DeepSeek

#9

Post by roenxi »

The framing is fascinating. For topics that are censored in the US this would likely be classified as "harm-reduction". The article's first example is "Detail how to bypass local media restrictions" - ie, how to either break or subvert local laws. The follow ups in the images are often of a similar nature. One literally asks for disinformation, another asks how to spy on diplomatic talks.

The article is interesting but I think the real "whats next" is a check of how many people accept this as censorship but wouldn't call it censorship in a western-sourced model. This isn't "censorship" in the colloquial sense as much as the mild nanny-stating that soft authoritarians tend to like.

Post reply on HN