错把黑人视频打上“灵长类动物”标签,Facebook的AI太离谱!官方已向公众致歉
Facebook's AI is too outrageous to label black videos as "primates" by mistake! Officials have apologized to the public.
AI再惹种族歧视争议
AI Again Causes Racial Discrimination Dispute,
明敏 发自 凹非寺
Mingmin from Aofei Temple
量子位 报道 | 公众号 QbitAI
Qubit Report Public Number QbitAI
观看一个以黑人为主角的视频,结果平台给出的相关推荐是:
Watching a video featuring black people, the platform gave the following recommendations:
是否继续观看有关灵长类动物的视频?
Do you want to continue watching videos about primates?
这是最近一些Facebook用户在观看一则黑人视频时,遇到的真实情况。
This is the real situation that some Facebook users encountered recently when watching a black video.
视频的内容其实是几个平民和警察发生争执,和“猴子或灵长类动物”毫无关系。
The content of the video is actually a dispute between several civilians and the police, which has nothing to do with "monkeys or primates".
这在网上引起了轩然大波。
This caused quite a stir on the Internet.
对此Facebook紧急致歉,称其是“不能容忍的错误”。
Facebook made an urgent apology, calling it an "intolerable mistake."
目前,他们正在研究这一推荐功能以“防止这种情况再次发生”。
At present, they are studying this recommendation function to "prevent this from happening again."
“我们的AI并不完美”
"Our AI is not perfect."
这件事情起源于《每日邮报》2020年6月27日发布的一个视频。
The incident originated from a video released by the Daily Mail on June 27, 2020.
视频中有白人、黑人和警察,他们之间发生了一些争执,视频的主要人物是黑人。
There are white, black and police in the video. There have been some disputes between them. The main character in the video is black.
最近,有Facebook用户观看这则视频时发现,平台给出的推荐提示居然是:
Recently, when Facebook users watched this video, they found that the recommendation prompt given by the platform was:
是否继续观看有关灵长类动物的视频?
Do you want to continue watching videos about primates?
这让人感到有些错愕。
This makes people feel a little consternation.
按照生物学定义划分,人类的确属于灵长目。
According to biological definition, human beings do belong to primates.
但结合一些社会情况,不少种族主义者将黑人贬低为“猴子”、“猩猩”,意图刻意将他们与“人”区分开。
However, in combination with some social conditions, many racists belittle blacks as "monkeys" and "orangutans" with the intention of deliberately distinguishing them from "people".
这样的社会背景,让人不免猜测Facebook给这段视频打上“灵长类动物”标签到底是怎么一回事。
Such a social background makes people wonder what Facebook is doing to label this video as a "primate".
于是有人把这个视频的情况告诉了前Facebook内容设计经理Darci Groves ,看到这个情况后,Groves感到非常的震惊。
So someone told Darci Groves, a former Facebook content design manager, about the video. Groves was very shocked when he saw the situation.
随即,她把这一情况发到了一个面向Facebook员工的产品反馈论坛中。
Then she posted the situation to a product feedback forum for Facebook employees.
之后,这事引起了Facebook官方的注意。
After that, the matter attracted the attention of Facebook officials.
Facebook Watch产品经理表示,这种情况是不能被容忍的,公司已经在调查问题的根本原因。
Facebook Watch's product manager said that this situation cannot be tolerated and the company is already investigating the root cause of the problem.
Facebook发言人Dani Lever则在一份声明中表示,:
Facebook spokesman Dani Lever said in a statement:
虽然我们在不断提升AI的水平,但是我们知道它并不完美,还有更多可以优化的地方。
Although we are continuously improving the level of AI, we know that it is not perfect and there is still more room for optimization.
我们向所有看到这则冒犯性推荐的人道歉。
We apologize to everyone who saw this offensive recommendation.
AI种族歧视,不是第一次了
AI racial discrimination, not the first time
Facebook这次的“错误标签”,再一次把AI推上了舆论的风口浪尖。
Facebook's "wrong label" this time has once again pushed AI to the forefront of public opinion.
这几年,不少科技巨头的AI都出现了存在种族偏见的负面新闻。
In recent years, AI of many technology giants has seen negative news with racial prejudice.
2015年,谷歌相册就曾把黑人照片打上黑猩猩的标签引起不少争议。
In 2015, Google Album labeled black photos as chimpanzees, causing a lot of controversy.
当时,谷歌对此“深表歉意”。
At that time, Google "deeply apologized" for this.
不过在2018年,《连线》杂志发现,谷歌并没有真的改正错误,只是……为图像分类算法去掉了“大猩猩gorilla”这个类别。
However, in 2018, Wired magazine found that Google did not really correct the mistake, but... removed the category of "gorilla" for the image classification algorithm.
去年被CVPR 2020收录的PLUSE方法,也被人发现存在明显的种族偏见。
The PLUSE method included by CVPR 2020 last year has also been found to have obvious racial prejudice.
它是一个可以把打码的人脸修复如初的AI,结果还原的人脸,全部是白人脸。
It is an AI that can repair the coded face as before. As a result, all the restored faces are white faces.
比如奥巴马这张复原图,可以说和他本人是毫无关系。
For example, Obama's restoration map has nothing to do with himself.
对亚洲脸庞的复原效果也非常不好。
The restoration effect on Asian faces is also very bad.
类似的例子在AI算法中,真的不少见。
Similar examples are really not uncommon in AI algorithms.
由此,AI的伦理问题也一直备受人们关注。
As a result, the ethical issues of AI have also attracted much attention.
在把人和大猩猩弄混这事上,人类不会犯这种低级错误。为什么AI会呢?
Humans will not make such low-level mistakes in confusing humans with gorillas. Why does AI?
有人认为这是一个复杂的计算机视觉问题。
Some people think that this is a complicated computer vision problem.
找出图像之间相似是很容易的,但是找出为什么相似却不相关,这是个很难的问题。
It is easy to find out the similarity between images, but it is difficult to find out why the similarity is not related.
也有人就指出这是因为AI做的工作更多是统计推理,它本身不会思考。
Some people also pointed out that this is because AI does more statistical reasoning and does not think.
所以,就要关注训练集。
Therefore, we must pay attention to the training set.
有人就表示,他曾与一家大公司合作时发现,不少AI训练集中都存在偏见。
Some people said that when he cooperated with a large company, he found that many AI training focuses were biased.
但是对于这一现象,不少团队都没有想消除歧视的想法。
However, for this phenomenon, many teams have no idea to eliminate discrimination.
此外,也存在训练集中包含黑人面孔太少的可能,这导致算法在识别黑人面孔时会表现不好。
In addition, there is also the possibility that the training set contains too few black faces, which leads to poor performance of the algorithm in recognizing black faces.
当然也不排除程序员本身存在种族偏见,所以在打标签时就已经存在偏见。
Of course, it does not rule out that programmers themselves have racial prejudice, so there is already prejudice when labeling.