近期,大量AI绘画席卷社交媒体。
这些照片看上去精致漂亮,完成度极高,但它们不来源于真正的艺术家,而是使用人工智能肖像应用程序制作的。
图源:今日邮报
这些照片由一款AI绘画应用制作,可以根据用户提交的照片,利用人工智能渲染出一幅数字肖像。
An app which uses artificial intelligence to render digital portraits based on photos users submit has gone viral worldwide.
该软件还会使用多种工具编辑人们脸上的所谓的瑕疵 (imperfection) ,从而使照片变得更美。
图源:Designboom
编辑照片的工具也多种多样,包括魔法触摸、替换或模糊背景、应用艺术滤镜和特效,或者为照片添加边框。甚至还可以将他们的照片转换成各种艺术风格,如动漫、卡通、素描或水彩画。
There are several tools for users to edit photos including magic touch, replace or blur out the background, apply art filters and special effects or add borders to their pictures. Users can also transfer their photos into various art styles such as anime, cartoons, sketches or watercolours.
图源:Standard
虽然这款应用在互联网上掀起了风暴 (take the internet by storm) ,但它也引发了围绕伦理、数据偏见等方面的争论。
为了开发AI绘画技术,开发人员和工程师会用大量数据训练模型,以识别和学习某些特征或风格。
一旦模型学会了这些信息,就可以把一张新图片以被训练好的风格重现出来。
Once the model learns that information, it can look at a new picture and reproduce that image in one of the styles it’s been trained to reproduce.
图源:NBC
艺术家们对这项技术背后隐含的道德伦理问题表示担忧。
Artists have raised concerns about the murky ethics underlying the technology.
因为AI绘画应用接受到的训练是基于艺术家在互联网上创作和发布的艺术作品。
一些艺术家声称,把用户支付的费用一小部分拿出来就可以让AI潦草绘制50张画作。这是对艺术家作品的贬低 (devalue) ,更有甚者,AI还可能会挪用 (appropriate) 创作者的作品,甚至包括他们的签名。
Some artists claim this not only devalues their own work, churning out 50 images at a fraction of the cost of a commission, but it is potentially appropriating their work, including their signature.
图源:Pinterest
“我们应该问:是否应该允许AI学习市面上的每一件艺术作品,让它模仿艺术家的独特风格?” 东北大学体验式人工智能研究所的伦理学负责人坎苏·坎卡说。
“We should ask: Should it be allowed to use every artwork out there to train AI models to imitate artists’ unique styles so closely?” says Cansu Canca , ethics lead for Northeastern’s Institute for Experiential AI .
霸王条款
用户可能只是想用这款软件绘制一张新的头像,但AI绘画软件的隐私政策允许该公司使用用户的“面部数据”来训练他们的人工智能算法。
Users might just want to use the app to create a new profile picture, but Lensa’s privacy policy allows the company to use their “face data” to train their AI algorithm.
该应用程序的使用条款还授予了该APP “永久的、不可撤销的、非排他性的、免版税的、全球性的、全额支付的、可转让的、可再授权的许可”,以使用在该APP中创建的肖像。
The app’s terms of use also give Lensa “perpetual, irrevocable, nonexclusive, royalty-free, worldwide, fully-paid, transferable, sub-licensable license” to use the portraits created in the app.
对于这些类似于霸王合同的使用条款,一位推特用户给出了她的建议。
图源:推特
1. You’re helping them train their Al.
首先,你在帮助他们训练Al。
2. They convert your photos to face data, which they claim to not retain but if you worry, you can request that they delete all your personal data by sending an email to: mailto:privacy@lensa-ai.com.
第二,他们将你的照片转换为面部数据,他们声称不会保留这些数据,但如果你担心的话,可以发送电子邮件到要求他们删除你的所有个人数据。
3. If you upload any of the images to social media, you’ve given them permission to use them in advertising. You can rescind permission by emailing: contact@lensa-ai.com.
第三,如果您将图像上传到社交媒体,则表示您已授予它们可以用来做广告,您可以通过电子邮件取消许可。
此外,一些女性用户也表示,这款应用会在没有提示的情况下创造过度性感的图像。
Some users, particularly women, have also said the app produces oversexualized images with no prompting.
图源:卫报
还有一些用户批评这款应用不会描绘亚洲面孔,甚至把他们的种族特征抹去了。
图源:JoySauce
“这些表现偏差更多的是人工智能训练出来的艺术风格而导致的,而非技术问题。”东北大学艺术与设计副教授詹妮弗说。
“These representation biases are more an indictment of the art styles the AI was trained on than the technology.” Jennifer Gradecki said, an associate professor of art and design at Northeastern.
“这些风格中有很多刻板印象,”詹妮弗说。“算法复制偏见,它们天生如此。 (“Algorithms replicate bias. They’re just built to do that.”) ”
图源:Pew Research Center
但也有专家表示,这不是重现偏见的借口。公司可以投入更多工作来清理和抓取输入算法的数据,把性别和种族等类别作为“控制信号”,使模型更加公平。
Companies can put in more work into scrubbing and scraping the data that is being fed to their algorithms, using categories like gender and race as “control signals” to make the model more fair.
但真正的问题在于,很多工程师和设计师都有意或无意地“没想到”这件事。