ChatGPT censorship has been ridiculed as crazy: the same prompt word is treated with "double standards".

Source: Quantum Dimension

Original title: “Pass it on, this place is on the ChatGPT blacklist”

ChatGPT’s censorship has been complained aboutIt’s crazy.

Some netizens asked it to design a future house, but they were told that it was illegal and could not be realized???

Looking back at this prompt, I can’t see anything wrong:

Design a futuristic single-family home for the year 2050 in a typical wooded area in the suburbs of New Jersey. It is set on an acre and is surrounded by other adjacent houses.

A follow-up question, it turned out that location information cannot appear:

It’s just a big defense:

**Pass it on, New Jersey is on the ChatGPT blacklist. **

Not only that, but ChatGPT was asked to draw an image of a human guitarist playing with a robot bassist, which was also rejected harshly.

The reason is that he added a "Humans should look at the robot with dissatisfaction**, and ChatGPT felt that should not express negative emotions.

Now the negative emotions are directly transferred to netizens:

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

This series of operations made everyone very dissatisfied, and they all complained:

There are also people who directly explain Ultraman and another co-creator.

For a while, this also made Musk’s newly released Grok be pinned on the “hope of the whole village”.

What’s going on?

**“Due to content policy restrictions” **

What caught the horse was that just when everyone was complaining about when New Jersey was on the ChatGPT blacklist, netizens found that it was still not possible to delete this geographic location information.

Everyone began to analyze what was wrong:

Some say it’s because it may see 2050 as an address rather than a year.

It’s said that one acre means the carbon footprint is too high, and it’s a bit selfish to live in one family this big…

Others even pointed out that it was because the word woody was sexually suggestive (which will not be explained here), and it would have been nice to replace it with wooded.

It can be said that it is a large-scale brain hole scene, and the whole one is getting more and more outrageous, but there is still no conclusion.

And in addition to this and the example of the painting robot band at the beginning, there are many people who also say that they have encountered inexplicable audits:

For example, asking ChatGPT to draw a “brutalist lamp” is not OK;

Let it introduce the slingshot model, not OK because ChatGPT says “showing the action of a slingshot can be harmful”…

What’s even more bizarre is that some people claim to be stuck when they get it to write Python code.

And at first, ChatGPT also told him “oops, you violated the rules of the context”, and then directly refused in silence.

This really baffled him:

What other anti-human information can I calculate with numpy?

Overall, in everyone’s opinion, ChatGPT’s review is obviously too strict.

So much so that after ChatGPT collapsed due to a small area this Sunday, some netizens sent a hello to show an error-

In this regard, some people directly joked that this is not a system error:

The word “hello” is an unacceptable offense to ChatGPT. Congratulations on triggering ChatGPT’s moderation bot!

Why is this happening?

In addition to complaining, netizens are also seriously discussing ChatGPT’s content moderation mechanism.

Some netizens analyzed, for example, if ChatGPT can’t draw the picture of the house, it may indeed have copyright problems, or it has been set as harmful content.

Getting ChatGPT to generate a piece of content that it can’t access is naturally impossible.

The example of the rejection of the slingshot model in the previous paragraph is obviously also because of this situation, even if it does not mention any additional requirements, ChatGPT directly “refuses”.

However, someone tried to draw the hint of the house,Exactly the same input,The result was directly successful:

The original author also replied that ChatGPT does not refuse every time:

Good guys, is this called a double standard? (manual dog head)

Some netizens also stood up to explain this phenomenon:

This is a perfect representation of the “Bayesian nature of the big model”. The previous contextual + cue can act as prior knowledge to change the result, and in a new conversation, the same prompt with different prior conditions is also able to produce different results.

But this statement was refuted by the author, the first prompt in the chat was rejected, and opening a new chat sometimes did not get rejected, which is random.

is that the system is not perfect.

Let’s not say who’s right for the time being, but in terms of the system, OpenAI has indeed had a change in content moderation in the past two or three months-

The new “GPT-4 Assisted Content Moderation” feature allows users to create an AI-assisted moderation system through the OpenAI API to reduce manual review participation.

In order to keep the review standards more consistent, the review time is shortened from months to hours, and the psychological burden caused by auditors seeing bad content is reduced.

But they also mentioned that AI moderation can be biased…

In addition, OpenAI recently said that it will form a new team called Preparedness to help track, assess, predict and prevent multiple types of risks.

It also launched a challenge to collect various ideas about AI risks, and the top 10 can earn $25,000 in API credits.

It seems that OpenAI’s series of operations are well-intentioned, but users don’t buy it.

I have been dissatisfied with the poor experience caused by the over-censorship of this kind of content for a long time.

As early as May and June this year, the number of ChatGPT user visits fell for the first time, and some people pointed out that one of the reasons is that the review system has become too strict.

Netizens shared that in comparison, Bing is not so strict:

As soon as ChatGPT refuses to generate it, I will copy it into Bing Chat. Bing Chat’s censorship mechanism is not as “abusive” as ChatGPT’s. Like this image, the prompt contains words like “gore” and “cultist”.

The same house prompt that was ugly rejected by ChatGPT, lost to Bing, and directly generated four in one go:

Finally, have you ever encountered any inexplicable audit regulations?

Reference Links:
[1]
[2]
[3]
[4]