论文部分内容阅读
The election season is winding up， and my social media is once again awash with political stories. Headlines stream： “Warren and Bernie’s awkward truce...”， “Trump sees his base growing...” and “The Fed’s real message...”. This is the America I see today.
The trouble is， it’s not the America you see or anyone else sees. It is my personally-curated version of reality. A constantly shifting mirage， evolving in real-time， depending on my likes and dislikes， what I click on， and what I share.
A recent Pew Research Center study found black social media users are more likely to see race-related news. The Mueller report suggests Russian efforts against Hillary Clinton targeted Bernie Sanders supporters. In October 2016， Brad Parscale， then President Trump’s 2016 digital director， told Bloomberg News that he targeted Facebook and media posts at possible Clinton supporters so that they would sit the election out.
Parscale―who， as of early August， has spent more （$9.2 million） on Facebook ads for Trump 2020 than the four top Democratic candidates combined―said that in 2016 he typically ran 50，000 ad variations each day， micro-targeting different segments of the electorate.
Algorithms are prejudiced
While political operatives exploiting yellow journalism is nothing new， the coupling of their manipulative techniques to a technologically-driven world is a substantial change. Algorithms are now the most powerful curators of information， whose actions enable such manipulation by creating our fractured informational multiverse.
And those algorithms are prejudiced. That may sound extreme， but let me explain.
In analyses conducted by myself and colleagues at University College London （UCL）， we modeled the behavior of social networks， using binary signals （1s and 0s） passed between simplified “agents” that represented people sharing of opinions about a divisive issue （say pro-life versus pro-choice or the merits of building a wall or not）.
Most “agents” in this model determine the signals they broadcast based on the signals they receive from those surrounding them （as we do sharing news and stories online）. But we added in a small number of agents we called “motivated reasoners，” who， regardless of what they hear， only broadcast their own pre-determined opinion.
Our results showed that in every case， motivated reasoners came to dominate the conversation， driving all other agents to fixed opinions， thus polarizing the network. This suggests that “echo chambers” are an inevitable consequence of social networks that include motivated reasoners. It goes deeper than you think： Two years after Charlottesville1， I’m fighting the conspiracy theory industrial complex.
So who are these motivated reasoners？ You might assume they are political campaigners， lobbyists or even just your most dogmatic Facebook friend. But， in reality， the most motivated reasoners online are the algorithms that curate our online news.
How technology generalizes
In the online media economy， the artificial intelligence in algorithms are single-minded in achieving their profit-driven agendas by ensuring the maximum frequency of human interaction by getting the user to click on an advertisement. But AIs are not only economically single-minded， they are also statistically simple-minded.
Take， for example， the 2016 story in The Guardian about Google searches for “unprofessional hair” returning images predominantly of black women.
Does this reveal a deep social bias towards racism and sexism？ To conclude this， one would have to believe that people are using the term “unprofessional hair” in close correlation with images of black women to such an extent as to suggest most people feel their hairstyles define “unprofessional.” Regardless of societal bias （which certainly exists）， this seems doubtful.
It isn’t all bad news for newspapers： I’m a journalism student in an era of closing newsrooms， ‘fake news.’ But I still want in.
Having worked in AI for 30 years， I know it is probably more statistically reliable for algorithms to recognize black women’s hairstyles than those of black men， white women， etc. This is simply an aspect of how algorithms “see，” by using overall features of color， shape， and size. Just as with real-world racism， resorting to simple features is easier for algorithms than deriving any real understanding of people. AIs codify this effect.
To be prejudiced means to pre-judge on simplified features， and then draw generalizations from those assumptions. This process is precisely what algorithms do technically. It is how they parse the incomprehensible “Big Data” from our online interactions into something digestible. AI engineers like me explicitly program generalization as a goal of the algorithms we design.
Given the simplifying features that algorithms use （gender， race， political persuasion， religion， age， etc.） and the statistical generalizations they draw， the real-life consequence is informational segregation， not unlike previous racial and social segregation. Dangerous， divisive consequences
Groups striving for economic and political power will inevitably exploit these divisions， using techniques such as targeted marketing and digital gerrymandering to categorize groups. The consequence is not merely the outcome in an election， but the propagation of deep divisions in the real world we inhabit.
Recently， Sen. Kamala Harris spoke about how federally-mandated desegregation busing transformed her life opportunities. Like her， I benefited from that conscious effort to mix segregated communities， when as a child in 1970s Birmingham， Alabama， black children were bused to my all white elementary school. Those first real interactions I had with children of a different race radically altered my perspective of the world.
It never gets easier： How many more birthdays will our journalist son， Austin Tice2， spend captive in Syria？
The busing of the past ought now inspire efforts to overcome the digital segregation we see today. Our studies at UCL indicate that the key to counteracting the natural tendency of algorithmically-mediated social networks to segregate is to technically promote mixing of ideas， through greater informational connectivity between people.
Practically， this may mean the regulation of online media， and an imperative for AI engineers to design algorithms around new principles that balance optimization with the promotion of diverse ideas. This scientific shift in perspective will ensure a healthier mix of information， particularly around polarizing issues， just like those buses enabled racial and social mixing in my youth.
選举季行将结束，我的社交媒体则再次充斥着政坛故事。新闻头条：“沃伦和伯尼的尴尬休战……”“特朗普喜看基本盘扩张……”“美联储的真实讯息……”。这就是我今天看到的美国。
问题是，这并非您看到的美国或任何其他人看到的美国。这是我的个人定制版现实。一幅不断移动的海市蜃楼图，根据我的赞和踩、点击及分享而实时演化。
皮尤研究中心最近的研究发现，黑人社交媒体用户更有可能看到种族相关的新闻。穆勒报告表明，俄罗斯人搞的反希拉里·克林顿动作的对象是伯尼·桑德斯的支持者。2016年10月，布拉德·帕斯卡尔——时任特朗普2016年总统竞选数字总监——向彭博新闻社透露，他以克林顿的潜在支持者为其在脸书和媒体帖子的受众，以便这些人选举时不去投票。
截至8月初，帕斯卡尔为特朗普2020年竞选在脸书广告上的花费（920万美元）比4名民主党支持率最高候选人的总和还多。他说，2016年，他通常每天投放5万个广告变体，精准定位不同的选民群体。
算法有偏见
尽管利用小报新闻的政治特工并非新鲜事物，但将他们的操纵技术与技术驱动的世界相结合却是实质巨变。算法乃现时最强大的信息管理员，通过创建破碎的信息多重宇宙，使这种操纵成为可能。
而且这些算法带有偏见。听起来可能有些极端，请容我解释。
我本人和伦敦大学学院的同事进行了多项分析。研究中，我们使用经简化的“代理人”之间传递的二进制信号（1和0）对社交网络的行为建模，这些信号代表人们就分歧问题发表意见（例如，生存优先还是选择优先，建墙不建墙到底哪个好）。
模型中，大多数“代理人”都是根据从周围人那里收到的信号来确定他们广播的信号（恰如我们在线分享新闻和故事时的行为）。但是，我们添加了少数称为“有动机的推理者”的代理人，他们无论听到什么都只会发表自己预设的意见。
我们的研究结果表明，在每种情况下，有动机的推理者最终都会主导对话，将所有其他代理人推向固定的观点，从而使网络两极化。这表明，只要社交网络存在有动机的推理者，“回声室”就是必然结果。 事情比您想得更深：夏洛茨维尔事件两年后，我还在与阴谋论产业复合体斗争。
那么，这些有动机的推理者究竟为何人？读者可能会认为是政治活动家、说客乃至其最自以为是的脸书好友。但实际上，网上有着最强动机的推理者是管理我们在线新闻的算法。
技术如何概括
在线媒体经济中，算法的人工智能一心一意通过让用户点击广告来确保最高频率的人机交互，从而实现其以利润为导向的议程。但是，人工智能不仅在经济上一心一意，在统计上也是一心一意。
以2016年《卫报》中有关谷歌搜索“不专业发型”的故事为例，反馈的图像主要是黑人女性。
这是否揭示出趋向种族主义和性别歧视的某种深层的社会偏见？要得出这个结论，必须得相信人们使用的“不专业的发型”一词与黑人女性的形象密切相关，以至暗示大多数人认为她们的发型定义了何为“不专业”。抛开社会偏见（确实存在）不谈，这似乎是可疑的。
对报纸而言，这并非全然是坏消息：我是一个生活在新闻编辑室日渐关闭（“假新闻”）时代的新闻专业学生。但我还是希望入局。
在人工智能领域工作了30年，我明白算法在识别黑人女性发型上可能要比识别黑人男性、白人女性等人群的发型在统计学上更为靠谱。这只是算法的一个方面，即使用颜色、形状和大小这些整体特征来“观看”。恰如现实世界中的种族主义，对于算法而言，诉诸简单特征要比真正理解人容易许多。人工智能将这种效应程序化。
带有偏见意味着基于简化的特征进行预判，并将此类假设进行概括。这个过程正是算法在技术上所做的。这是他们将在线交流中无法理解的“大数据”解读为可消化内容的过程。对我这样的人工智能工程师而言，很明确，将这种概括设定为我们所设计的算法的一个目标。
鉴于算法使用的简化特征（性别、种族、政治立场、宗教、年龄等）以及它们得出的统计概括，现实生活所受到的影响就是信息隔离，与以往的种族隔离和社会隔离并无二致。
危险且分裂性的后果
旨在攫取经济和政治权力的团体将无可避免地利用這种细分，使用定向营销和不正当的数字划分等技术将团体归类。这种做法不仅影响个别选举的结果，还在我们所处的现实世界中散播深层次的分裂。
参议员卡玛拉·哈里斯近期曾谈及联邦政府强制实行的废除种族隔离校车制度如何改变了她的人生机遇。笔者儿时生活在1970年代的亚拉巴马州伯明翰，当黑人儿童乘校车来到我所在的全白人小学时，和哈里斯一样，我也从有意识消除种族隔离社区的努力中得益。那些与来自另一种族的孩子们最初的真正互动，从根本上刷新了我的世界观。
事情从来就不容易：不知道我们的记者小伙儿奥斯汀·蒂斯还有多少个生日将在叙利亚的囚禁中度过？
过往的废除种族隔离校车制度理当激发我们现在去克服今日所见的数字隔离。我们在伦敦大学学院的研究表明，要抵抗算法调制的社交网络隔离的自然趋势，关键在于通过人与人之间更强的信息互联来从技术上促进观念的交融。
实际上，这可能意味着对在线媒体的监管，以及要求人工智能工程师围绕新原则设计算法，这些原则应当在最优结果与多元观念推广之间达致平衡。科学转变视角将确保更健康的信息融合，尤其事关两极分化的问题，恰如那些在我青少年时代实现了种族和社会融合的校车一样。
The trouble is， it’s not the America you see or anyone else sees. It is my personally-curated version of reality. A constantly shifting mirage， evolving in real-time， depending on my likes and dislikes， what I click on， and what I share.
A recent Pew Research Center study found black social media users are more likely to see race-related news. The Mueller report suggests Russian efforts against Hillary Clinton targeted Bernie Sanders supporters. In October 2016， Brad Parscale， then President Trump’s 2016 digital director， told Bloomberg News that he targeted Facebook and media posts at possible Clinton supporters so that they would sit the election out.
Parscale―who， as of early August， has spent more （$9.2 million） on Facebook ads for Trump 2020 than the four top Democratic candidates combined―said that in 2016 he typically ran 50，000 ad variations each day， micro-targeting different segments of the electorate.
Algorithms are prejudiced
While political operatives exploiting yellow journalism is nothing new， the coupling of their manipulative techniques to a technologically-driven world is a substantial change. Algorithms are now the most powerful curators of information， whose actions enable such manipulation by creating our fractured informational multiverse.
And those algorithms are prejudiced. That may sound extreme， but let me explain.
In analyses conducted by myself and colleagues at University College London （UCL）， we modeled the behavior of social networks， using binary signals （1s and 0s） passed between simplified “agents” that represented people sharing of opinions about a divisive issue （say pro-life versus pro-choice or the merits of building a wall or not）.
Most “agents” in this model determine the signals they broadcast based on the signals they receive from those surrounding them （as we do sharing news and stories online）. But we added in a small number of agents we called “motivated reasoners，” who， regardless of what they hear， only broadcast their own pre-determined opinion.
Our results showed that in every case， motivated reasoners came to dominate the conversation， driving all other agents to fixed opinions， thus polarizing the network. This suggests that “echo chambers” are an inevitable consequence of social networks that include motivated reasoners. It goes deeper than you think： Two years after Charlottesville1， I’m fighting the conspiracy theory industrial complex.
So who are these motivated reasoners？ You might assume they are political campaigners， lobbyists or even just your most dogmatic Facebook friend. But， in reality， the most motivated reasoners online are the algorithms that curate our online news.
How technology generalizes
In the online media economy， the artificial intelligence in algorithms are single-minded in achieving their profit-driven agendas by ensuring the maximum frequency of human interaction by getting the user to click on an advertisement. But AIs are not only economically single-minded， they are also statistically simple-minded.
Take， for example， the 2016 story in The Guardian about Google searches for “unprofessional hair” returning images predominantly of black women.
Does this reveal a deep social bias towards racism and sexism？ To conclude this， one would have to believe that people are using the term “unprofessional hair” in close correlation with images of black women to such an extent as to suggest most people feel their hairstyles define “unprofessional.” Regardless of societal bias （which certainly exists）， this seems doubtful.
It isn’t all bad news for newspapers： I’m a journalism student in an era of closing newsrooms， ‘fake news.’ But I still want in.
Having worked in AI for 30 years， I know it is probably more statistically reliable for algorithms to recognize black women’s hairstyles than those of black men， white women， etc. This is simply an aspect of how algorithms “see，” by using overall features of color， shape， and size. Just as with real-world racism， resorting to simple features is easier for algorithms than deriving any real understanding of people. AIs codify this effect.
To be prejudiced means to pre-judge on simplified features， and then draw generalizations from those assumptions. This process is precisely what algorithms do technically. It is how they parse the incomprehensible “Big Data” from our online interactions into something digestible. AI engineers like me explicitly program generalization as a goal of the algorithms we design.
Given the simplifying features that algorithms use （gender， race， political persuasion， religion， age， etc.） and the statistical generalizations they draw， the real-life consequence is informational segregation， not unlike previous racial and social segregation. Dangerous， divisive consequences
Groups striving for economic and political power will inevitably exploit these divisions， using techniques such as targeted marketing and digital gerrymandering to categorize groups. The consequence is not merely the outcome in an election， but the propagation of deep divisions in the real world we inhabit.
Recently， Sen. Kamala Harris spoke about how federally-mandated desegregation busing transformed her life opportunities. Like her， I benefited from that conscious effort to mix segregated communities， when as a child in 1970s Birmingham， Alabama， black children were bused to my all white elementary school. Those first real interactions I had with children of a different race radically altered my perspective of the world.
It never gets easier： How many more birthdays will our journalist son， Austin Tice2， spend captive in Syria？
The busing of the past ought now inspire efforts to overcome the digital segregation we see today. Our studies at UCL indicate that the key to counteracting the natural tendency of algorithmically-mediated social networks to segregate is to technically promote mixing of ideas， through greater informational connectivity between people.
Practically， this may mean the regulation of online media， and an imperative for AI engineers to design algorithms around new principles that balance optimization with the promotion of diverse ideas. This scientific shift in perspective will ensure a healthier mix of information， particularly around polarizing issues， just like those buses enabled racial and social mixing in my youth.
選举季行将结束，我的社交媒体则再次充斥着政坛故事。新闻头条：“沃伦和伯尼的尴尬休战……”“特朗普喜看基本盘扩张……”“美联储的真实讯息……”。这就是我今天看到的美国。
问题是，这并非您看到的美国或任何其他人看到的美国。这是我的个人定制版现实。一幅不断移动的海市蜃楼图，根据我的赞和踩、点击及分享而实时演化。
皮尤研究中心最近的研究发现，黑人社交媒体用户更有可能看到种族相关的新闻。穆勒报告表明，俄罗斯人搞的反希拉里·克林顿动作的对象是伯尼·桑德斯的支持者。2016年10月，布拉德·帕斯卡尔——时任特朗普2016年总统竞选数字总监——向彭博新闻社透露，他以克林顿的潜在支持者为其在脸书和媒体帖子的受众，以便这些人选举时不去投票。
截至8月初，帕斯卡尔为特朗普2020年竞选在脸书广告上的花费（920万美元）比4名民主党支持率最高候选人的总和还多。他说，2016年，他通常每天投放5万个广告变体，精准定位不同的选民群体。
算法有偏见
尽管利用小报新闻的政治特工并非新鲜事物，但将他们的操纵技术与技术驱动的世界相结合却是实质巨变。算法乃现时最强大的信息管理员，通过创建破碎的信息多重宇宙，使这种操纵成为可能。
而且这些算法带有偏见。听起来可能有些极端，请容我解释。
我本人和伦敦大学学院的同事进行了多项分析。研究中，我们使用经简化的“代理人”之间传递的二进制信号（1和0）对社交网络的行为建模，这些信号代表人们就分歧问题发表意见（例如，生存优先还是选择优先，建墙不建墙到底哪个好）。
模型中，大多数“代理人”都是根据从周围人那里收到的信号来确定他们广播的信号（恰如我们在线分享新闻和故事时的行为）。但是，我们添加了少数称为“有动机的推理者”的代理人，他们无论听到什么都只会发表自己预设的意见。
我们的研究结果表明，在每种情况下，有动机的推理者最终都会主导对话，将所有其他代理人推向固定的观点，从而使网络两极化。这表明，只要社交网络存在有动机的推理者，“回声室”就是必然结果。 事情比您想得更深：夏洛茨维尔事件两年后，我还在与阴谋论产业复合体斗争。
那么，这些有动机的推理者究竟为何人？读者可能会认为是政治活动家、说客乃至其最自以为是的脸书好友。但实际上，网上有着最强动机的推理者是管理我们在线新闻的算法。
技术如何概括
在线媒体经济中，算法的人工智能一心一意通过让用户点击广告来确保最高频率的人机交互，从而实现其以利润为导向的议程。但是，人工智能不仅在经济上一心一意，在统计上也是一心一意。
以2016年《卫报》中有关谷歌搜索“不专业发型”的故事为例，反馈的图像主要是黑人女性。
这是否揭示出趋向种族主义和性别歧视的某种深层的社会偏见？要得出这个结论，必须得相信人们使用的“不专业的发型”一词与黑人女性的形象密切相关，以至暗示大多数人认为她们的发型定义了何为“不专业”。抛开社会偏见（确实存在）不谈，这似乎是可疑的。
对报纸而言，这并非全然是坏消息：我是一个生活在新闻编辑室日渐关闭（“假新闻”）时代的新闻专业学生。但我还是希望入局。
在人工智能领域工作了30年，我明白算法在识别黑人女性发型上可能要比识别黑人男性、白人女性等人群的发型在统计学上更为靠谱。这只是算法的一个方面，即使用颜色、形状和大小这些整体特征来“观看”。恰如现实世界中的种族主义，对于算法而言，诉诸简单特征要比真正理解人容易许多。人工智能将这种效应程序化。
带有偏见意味着基于简化的特征进行预判，并将此类假设进行概括。这个过程正是算法在技术上所做的。这是他们将在线交流中无法理解的“大数据”解读为可消化内容的过程。对我这样的人工智能工程师而言，很明确，将这种概括设定为我们所设计的算法的一个目标。
鉴于算法使用的简化特征（性别、种族、政治立场、宗教、年龄等）以及它们得出的统计概括，现实生活所受到的影响就是信息隔离，与以往的种族隔离和社会隔离并无二致。
危险且分裂性的后果
旨在攫取经济和政治权力的团体将无可避免地利用這种细分，使用定向营销和不正当的数字划分等技术将团体归类。这种做法不仅影响个别选举的结果，还在我们所处的现实世界中散播深层次的分裂。
参议员卡玛拉·哈里斯近期曾谈及联邦政府强制实行的废除种族隔离校车制度如何改变了她的人生机遇。笔者儿时生活在1970年代的亚拉巴马州伯明翰，当黑人儿童乘校车来到我所在的全白人小学时，和哈里斯一样，我也从有意识消除种族隔离社区的努力中得益。那些与来自另一种族的孩子们最初的真正互动，从根本上刷新了我的世界观。
事情从来就不容易：不知道我们的记者小伙儿奥斯汀·蒂斯还有多少个生日将在叙利亚的囚禁中度过？
过往的废除种族隔离校车制度理当激发我们现在去克服今日所见的数字隔离。我们在伦敦大学学院的研究表明，要抵抗算法调制的社交网络隔离的自然趋势，关键在于通过人与人之间更强的信息互联来从技术上促进观念的交融。
实际上，这可能意味着对在线媒体的监管，以及要求人工智能工程师围绕新原则设计算法，这些原则应当在最优结果与多元观念推广之间达致平衡。科学转变视角将确保更健康的信息融合，尤其事关两极分化的问题，恰如那些在我青少年时代实现了种族和社会融合的校车一样。