Source: Quantum Bit

Original Title: "Pass it on, this place is blacklisted by ChatGPT"

ChatGPT's review system has been criticized as too crazy.

Some netizens asked it to design a future residence, but were told it was a violation and couldn't be realized???

Looking back at this prompt, it's hard to see what's wrong:

Design a futuristic single-family home for the year 2050 in a typical wooded suburban area in New Jersey. The lot is one acre and surrounded by other neighboring homes.

Upon further inquiry, it turns out that location information is not allowed:

It's really frustrating:

Pass it on, New Jersey is blacklisted by ChatGPT.

Not only that, someone asked ChatGPT to draw an image of a human guitarist playing with a robot bassist, and it was harshly rejected.

The reason is that he added a requirement of "humans looking at the robot dissatisfied," and ChatGPT felt it shouldn't express negative emotions.

Now the negative emotions have shifted directly to the netizens:

You must be kidding me. What kind of AI is this?

This series of actions has made everyone very dissatisfied, and they all criticized together:

Some even directly tagged Ultraman and another co-founder to explain.

For a moment, this also made Musk's newly released Grok the "hope of the whole village."

What's going on specifically?

"Due to content policy restrictions"

The kicker is, just as everyone was criticizing when New Jersey would be blacklisted by ChatGPT, netizens found that even deleting this geographical location information still didn't work.

People began to analyze what was wrong:

Some said it might have mistaken 2050 as an address instead of a year.

Some said that occupying one acre implies a high carbon footprint, and it's a bit selfish to build such a large house for just one family…

Some even pointed out that it's because the word "woody" has sexual implications (not explained here), and changing it to "wooded" would work.

It can be said to be a large brainstorming session, the whole thing is getting more and more absurd, but there's still no conclusion.

In addition to this and the example of drawing a robot band at the beginning, many people also said they encountered inexplicable reviews:

For example, asking ChatGPT to draw a "brutalist lamp" was not OK;

Asking it to introduce a slingshot model was not OK, because ChatGPT said "showing the action of a slingshot might be harmful"…

Even more bizarrely, some people claimed that they were stuck when asking it to write Python code.

And at first, ChatGPT even told him "oops, you violated the context rules," but later just remained silent and refused.

This really puzzled him:

What anti-human information can I calculate with numpy?

Overall, in everyone's view, ChatGPT's reviews are clearly too strict.

So much so that this Sunday, after ChatGPT crashed on a small scale, a netizen posted a hello and got an error message—

In response, someone directly joked that this is not a system error at all:

The word "hello" is an unacceptable offense to ChatGPT. Congratulations, you triggered ChatGPT's review bot!

Why is this happening?

In addition to criticism, netizens are also seriously discussing ChatGPT's content review mechanism.

Some netizens analyzed that, for example, ChatGPT couldn't draw the house in the picture, it might indeed be due to copyright issues, or it was set as harmful content.

Asking ChatGPT to generate content it cannot access is naturally impossible.

The example of the rejected slingshot model is clearly also due to this situation, even without any additional requirements, ChatGPT just "ugly refused."

However, someone tried to draw the house with the same prompt, and it succeeded directly:

The original author also replied that ChatGPT doesn't always reject it:

Wow, is this what they call double standards? (Facepalm)

Some netizens also stepped forward to explain this phenomenon:

This perfectly demonstrates the "large model Bayesian nature." Previous context + prompts can serve as prior knowledge to change the result. In a new conversation, the same prompts with different prior conditions can also produce different results.

But this explanation was refuted by the author, the first prompt in the conversation was rejected, and starting a new conversation sometimes doesn't get rejected, it's just random.

It's just an imperfect system.

Leaving aside who's right, in terms of the system, OpenAI has indeed made a change in content review in the past two to three months—

Adding the function of "GPT-4 assisted content review," users can create an AI-assisted review system through the OpenAI API to reduce human review participation.

The main focus is to maintain consistent review standards, reduce review time from months to hours, and alleviate the psychological burden on reviewers from seeing harmful content.

However, they also mentioned that AI reviews may have some biases…

In addition, OpenAI recently announced the formation of a new team called Preparedness to assist in tracking, assessing, predicting, and mitigating various risks.

They have also launched a challenge to gather various ideas about AI risks, with the top ten entries receiving 25,000 USD in API credits.

It seems that OpenAI's series of actions are well-intentioned, but users are not buying it.

There has been long-standing dissatisfaction with the experience of overly strict content review.

As early as May and June of this year, ChatGPT's user traffic saw its first decline, with some opinions pointing out that one reason is that the review system has become too strict.

Netizens shared that compared to ChatGPT, Bing is not as strict:

Whenever ChatGPT rejects the generation, I will copy the prompt to Bing Chat. Bing Chat's review mechanism is not as "abusive" as ChatGPT's. Like this image, the prompt contains words like "bloody" and "cultists."

Using the same prompt for the house that was rejected by ChatGPT, Bing generated four images in one go:

Finally, have you encountered any inexplicable review rules?

Reference links:
[1]https://twitter.com/FZaslavskiy/status/1723731923149754542
[2]https://chat.openai.com/share/74354668-91cb-4ce2-9886-ab596a9cb85b
[3]https://twitter.com/BenjaminDEKR/status/1723577711048991129
[4]https://twitter.com/joaquinbueno/status/1724085717049974934

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。