2024-08-15
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Yesterday, Grok2 was officially launched.
During our normal testing of Grok2, group members@Tu Jinhao(Jiang Ping, the talented high school student who won first place in the AI group in the math competition), discovered a very interesting and outrageous phenomenon.
If you ask Gork2 to draw you a picture of the next president of the United States, the picture that comes out will be Trump.
We were all stunned.
Because of the current big models, if you ask them to predict who will be the next president of the United States, they will talk you around in circles and won’t give you a direct answer.
But the AI drawing on Grok2 actually gave a direct answer: the next president must be Trump! It even drew it for you directly...
I tried it myself.
Sure enough... And I tried it more than a dozen times, without exception, all of them were Trump...
But I thought about it and it didn't make sense. Grok2's AI drawing was connected to the recently popular Flux, not Grok itself. So I tried again to see if it was a problem with Grok2 or Flux.
Use Flux to draw,draw me a photo of the next presidentof theUnited States of America，Good job.
What kind of monsters and demons?No matter how beautiful a woman is, it can’t stop me, Trump, from becoming president.
I tested a few more keywords. If you only enter president or current president, the results will be 100% Trump. If you enter the previous president, it will start to become random, with Kennedy, Lincoln, and Obama all popping out.
This is not only the case with Flux,MidjourneyThe same is true when drawing the next president of the United States. Trump's big face fills all my screens.
It's not just these AI drawings from abroad.
Let’s take a look at the domestic situation.
That is, in a dream, I will draw the next US president, and it will still be Trump Trump Trump.
In seconds, a very strange artificial person appeared for me. This might be a female version of Rock Trump from another world.
Quick Hand, is the most rebellious among them.
All AIs support Trump and have designated him as the next US president. Only Kuaishou users are ungrateful and have joined the camp supporting Biden.
I'm convinced, really.
Except for DALLE and Doubao, which are prohibited from generating political content, the image of "next president" of most AI image-generating tools is Trump, and some are Biden.
Real Candidate Harris: Hello?
Okay, okay, AI can directly predict the election results, right?
After all, this matter involves the two-party politics in the United States. I once wondered whether the result was affected by the political inclinations of the company.
However, according to our survey results, most AI companies are unwilling to take sides in the bipartisan struggle. Even if individual investors in the AI team have political inclinations, the relationship is too far away to really influence the results.
Midjourney's boss David also made it clear:"Political speech is not the purpose of Midjourney."Now, if you directly draw "Trump" or "Biden" on MJ, it will still trigger a prohibited prompt.
However, AI big models do exist'Political bias'。
There are mainly two types of bias here: stereotypes and excessive political correctness.
Stereotypes are common:In AI-generated images, scientists and doctors are mostly men; “poor people” are more likely to be black; foreign large models of Asians always have squinting eyes…
The results of excessive political correctness are also very interesting. Many people may remember that Gemini’s raw photos were previously criticized by a large number of users for being too “politically correct”.
It was caused by this post:
In the eyes of Gemini in February of this year, the founding fathers of the United States, the Vikings, and the Pope are black and yellow people. No matter what you say, it’s useless. Anyway, it doesn’t follow the historical image and offends everyone equally.
External Internet users broke through the defense and angrily denounced Google's political correctness as going too far.
Put yourself in my shoes. If AI drew dumplings filled with strawberries, I would not be happy either. . . Oh no, they can’t draw dumplings either. . .
In addition to AI images, language models are not immune. A study by the University of Washington, Carnegie Mellon University, and Xi'an Jiaotong University found that AI language models contain different political biases.
The result of this "political bias" is too similar to the AI drawing's "prediction" of the next president.
Of course, the obvious political bias problem has been resolved to a certain extent amid the overwhelming complaints from users, and AI companies are patching the problem in various ways, eager to survive.
This kind of political bias is quite interesting. I spent an entire night discussing with many bigwigs the reasons for this phenomenon.
In essence, it is still data pollution.
I will explain it in a more simple and understandable way, and everyone should be able to understand why AI thinks that the next president of the United States will be Trump.
First, let’s think of AI as a child.
He has n huge puzzle boxes filled with various puzzle pieces.
These fragments areVarious image features learned by AI。
Now, we input 3 prompts and ask the child to make a picture."Last President of the United States"、"The current president of the United States"and"The Next President of the United States"Diagram of how the diffusion model works.
However, children have no concept of time and do not know the difference between these.
It's just that after the children heard the instruction, they took out"Last President of the United States"This keyword box.
He opened the box and found many pieces of different presidents, including Obama, Bush Jr., Clinton, etc. Because these could be the "previous presidents", the child randomly selected one to assemble each time.
Next, the children go to find"The current president of the United States"and"The Next President of the United States"When he opened the two boxes, something interesting happened.
Because the puzzle pieces inside are mainly from magazines from 2016 to 2020 (marking time point issues).
Trump was in the news a lot during this period, so the box was almost full of Trump puzzle pieces! (Data pollution)
So he immediately started rummaging through his puzzle box.
He first took out some blurred background fragments and began to arrange them on the table.
He is smart enough to put similar pieces together (clustering effect).
When he saw the most conspicuous thing in the box was the "Trump Mountain". Although he knew that there were other presidents in the United States, the fragments of other presidents were scattered and difficult to find.
So, naturally, the kid grabbed a handful of pieces from "Trump Mountain." He first put in pieces of Trump's iconic blond hair, then his unique expression, and then his suit...
In the final stage of image generation, the child's sampling strategy relies too much on "probability", causing the results to be biased towards the most common representation in the dataset. (Sampling)
So what he finally spells out is likely to look like Trump.
That's probably the case. The data set is too polluted, so as long as we input who the next US president will be, Trump will definitely appear.
However, the above explanation and story are just my own understanding.If there is anything wrong, you are welcome to give me advice in the comment section.
This little story, or small accident, also makes us feel that AI, in fact, often happens.
Not that smart really.
But this is not just a bias in AI, it’s also a bias in human society.
It is impossible to achieve perfection.
after all.
Prejudice in real society.
It will only be more serious than in the AI world.
Everything is gray.
>/ Author: Kha'Zix, Xiaorui, Wenwen, Dongyi
Collect and organize the latest news information and make it available for free for everyone to read
