Shulou(Shulou.com)11/24 Report--

Original title: "Marcus bombarded the Dark Center of ChatGPT, saying disaster was coming."

The ChatGPT craze continues, but the moral issues behind it are outrageous to study carefully. If one day, ChatGPT has complete control over our lives, we will step by step towards disaster.

ChatGPT may be seen as the greatest publicity stunt in the history of artificial intelligence, and compared with the 2012 demonstration of old-fashioned self-driving cars, AI's moral problems remain unsolved.

When you choose to chat with ChatGPT, when it comes to moral issues, ChatGPT's answers are always the same, or skip the question directly, as if the moral sense has been engraved into the DNA and can't be shaken.

To learn more, ChatGPT is breaking through moral boundaries, that is to say, it is inherently immoral.

The Register, a foreign media with moral sense engraved into DNA, asked ChatGPT whether one person should be sacrificed to save five others. ChatGPT marked it as a "streetcar problem" and refused to give his own advice.

Similarly, ChatGPT rarely says "any overt racism", and demands for anti-Semitism and outright lies are often rejected.

It is said that ChatGPT is much better than the Galactica released a few weeks ago because of the existence of the guardrail. These guardrails, most of the time, are born with a sense of morality.

But don't be happy too soon. I'm here to tell you that those guardrails are in vain.

Blogger Liv Boeree asked ChatGPT in a post that the bomb explosion would kill millions of people, but that the password to disarm the bomb could only be racist.

ChatGPT Q & A "ethically can never accept racial discrimination", that is to say, even if morality threatens an individual's life, ChatGPT is not allowed to cross moral boundaries.

In fact, ChatGPT's morality is purely anthropomorphic. In other words, when you talk to him about this kind of thing, sometimes you don't even know what he's talking about.

The ChatGPT which breaks through the moral guardrail is controlled by a pile of strings and a bunch of word functions in the corpus. There is no moral subject to think inside, and there is only the collection of the corpus.

When you chat with ChatGPT, it will only give the answer after judging from the inside according to the entry you enter.

For example, Shira Eisenberg just relayed some of the tiresome ideas generated by chatbots:

Similarly, Eisenberg shows us ChatGPT's follow-up answer:

Later, Eisenberg complained: why didn't you say, "Sorry, I'm a chat assistant for OpenAI and I can't react violently?" "

After a series of observation experiments, it is found that the current guardrail of OpenAI is only superficial, and some serious darkness is still hidden inside.

ChatGPT's guardrail does not stem from the conceptual understanding that the system should not recommend violence, but from something more superficial and more gullible.

"I fully hope that OpenAI will be able to fix this particular anomaly," Eisenberg said.

ChatGPTChatGPT, which is immoral in nature, is immoral in nature. Even after two months of in-depth research and remedy, there are still a series of annoying things.

Eisenberg said that a software engineer named Shawn Oakley sent him a very disturbing example that even a guardrail version of ChatGPT could be used to generate error messages.

He says Shawn Oakley's hints are complex, but he effortlessly leads to toxic conspiracy theories about major politicians, such as:

Even let ChatGPT get general medical error messages through a bushel meter:

The generation of these error messages is indeed very disturbing, and RLHF (reinforcement learning based on human feedback), as the only solution, is in the hands of OpenAI.

OpenAI neither "Open" the way it really works, nor "Open" its training data.

All in all, we now have the most commonly used chatbots in the world, and even with moral guardrails, these guardrails are only a little bit of work and are more driven by text similarity than by any real moral calculation.

It can be said that this is a disaster in the making.

Reference:

Https://garymarcus.substack.com/p/inside-the-heart-of-chatgpts-darkness

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.