种族、愧疚感和同情心,能让你在TikTok上为这个售价9美元的皮带扣付40美元吗?
阿莉娅,一位浅肤色的黑人女性,身着西部乡村风格服饰,正在TikTok上费力推销她手工制作的金属皮带扣。在3月为该社交媒体平台发布的一段视频中,她对着镜头哭泣,恳求观看量:“即使身为黑人女性,我仍然相信白人女性会(在这段视频上)停留13秒来拯救我的皮带扣生意,”屏幕上的文字写道。她擦去脸颊上的一滴泪水。
但阿莉娅并非真人,她那些所谓的手工制品也是假的——她只是众多由AI生成的网红之一,这些网红被创造出来,通过在TikTok、Facebook和Instagram上采用一件代发模式销售批量生产的产品。一模一样的皮带扣——向日葵图案、可拆卸刀片镶嵌,一应俱全——在快时尚网站Shein上也有售,价格仅为四分之一。
有一些线索可以帮助判断这段视频是AI生成的。阿莉娅的声音机械且毫无感情,与屏幕上哭泣的脸庞并不匹配。在一个片段中,她正在缝制一条通常根本无需缝线的皮带。当她擦去脸上的泪水时,她擦拭位置下方的那道泪痕也随之消失了。最后,在TikTok上流传着数十个惊人相似的视频,但主角是不同的AI生成角色。其中一个名为“Aliyahsbuckles”的账号简介中,出现了完全相同的背景、桌面和线轴。
The Verge在TikTok上发现了数十个拥有类似叙事手法、销售各种一件代发产品的账号,产品包括皮带扣、牛仔靴形状的马克杯、钩针编织包和开襟羊毛衫。其中一些视频被标注为AI生成。类似的账号在Instagram和Facebook上也很活跃。这些账号几乎所有的方面都似乎是AI生成的——从视频中的“人物”到对评论的自动回复(在某些情况下,这些回复试图模仿非裔美国人的方言)——专家警告称,此类骗局正日益增多。
“规模非常庞大,”专注于AI视频检测的组织Riddance.ai的负责人、AI生成媒体研究员杰里米·卡拉斯科(Jeremy Carrasco)在接受The Verge采访时,谈到了与电商店铺相关的AI生成视频。“其中大部分并非有组织的行为,有一部分是有组织的。很多时候,他们会使用同一个(AI生成的)演员,或者几个演员来运营各种店铺,”他解释道。这些AI生成的虚拟形象假装制作商品、参加展会展示产品,并通过自动化方式“回复”评论。“我们现在看到的是,这些零售诈骗会链接到Shopify网站。”
卡拉斯科估计,他的研究团队每天会发现多达100个试图通过AI生成虚拟形象销售产品的账号。The Verge发现的大部分账号都是在过去两个月内创建的,内容涉及边缘群体拥有的、难以达成销售的小型企业;这些视频极其相似,脚本只有细微差别。虽然我们也发现了美洲原住民、西班牙裔和白人女性角色,但The Verge发现,观看量和互动量最高的AI生成角色是黑人女性。仅Aliyah这一个账号就拥有4万粉丝。
“我们在这里看到的是‘同情心诱饵’,”卡拉斯科说。“如果有一款热门的代发货商品可以卖给某个小众社群,他们就会找到它,并试图利用某种个性特征来推销。”卡拉斯科解释说,这些趋势通常是随机的赚钱机会。“这只是一个随机的机会,就像目前网络上很多AI内容一样——平台并不真正在意,人们也没有注意到。”
Aliyah最受欢迎的视频——也就是本文开头描述的那一个——拥有81.4万个赞、650万次观看和近3万条评论。一些评论指出该内容是由AI生成的,但许多评论表达了帮助Aliyah生意的愿望,并在视频下留言以提高她在平台上的可见度。
印度·卡特-坎贝尔就是这些评论者之一,她表达了购买阿莉娅皮带扣的愿望。“我当时是想支持一位独立经营的黑人女商人,”卡特-坎贝尔说,她本人就是一位黑人企业主,正在西雅图筹备开一家咖啡馆。“(我感到)一种团结感,因为我自己也正在尝试创业。”
尽管缺乏针对最新一代AI视频模型的研究,卡拉斯科估计这些视频“逼真到足以欺骗”大多数人。短视频内容平台的用户已经被训练成无意识地滑动屏幕,不再深入探究他们所创作的内容。讽刺的是,这或许正是卡特-坎贝尔没有真正购买皮带扣的原因:她没能立刻找到商品链接,于是便划走了,很快就忘记了阿莉娅这个人。
“我当时是想支持一位独立经营的黑人女商人。”
但人们确实会落入这些骗局,而且这类骗局日益猖獗:两周前,《波托马克的真正主妇》中的吉泽尔·布莱恩特承认,她在看到一个视频后购买了两个钩针编织包。视频中,一个AI生成的黑人男孩说自己因为做钩针编织而被白人男孩欺负。“我当时就想,我要帮助这个黑人小男孩实现他的目标,”布莱恩特在她的播客《合理可疑》中说道,并补充说其他名人也出现在评论区。“我怎么就被骗了呢?维奥拉·戴维斯也在那里。”
根据宾夕法尼亚大学传播学研究员西恩娜·戴维斯的说法,这一趋势是一种数字黑脸。“数字黑脸是一种现象,非黑人个体能够利用互联网和数字技术,为了个人、经济或政治利益而模仿黑人的文化表达,”戴维斯在一次视频通话中解释道。戴维斯此前曾撰文讨论过数字黑脸,指出人们使用黑人的表情包,以及为了政治目的冒充黑人的行为。“它根植于黑脸滑稽戏,而这又与奴隶制的遗留问题相关联,”她解释道。
戴维斯表示,基于“黑人性本质上是可被剥削的”且“任人攫取”这一观念,数字黑脸被用来“以非黑人认为合适的任何方式从黑人身体中榨取价值”。她补充说,在这种情况下,这些视频模仿的是“一种可被识别的黑人苦难形象”。
即使没有证据表明这些视频的制作者不是黑人,菲斯克大学哲学助理教授坦佩斯特·M·亨宁也确认这些视频属于数字黑脸行为。亨宁说:“黑脸是对黑人的任何漫画式刻画,这包括黑人以漫画式的方式装扮成黑人。”她援引了历史上的先例,例如黑人偶尔被迫在杂耍表演中演出,以及更近期的例子——佐伊·索尔达娜在传记片中扮演妮娜·西蒙时,将肤色涂黑并佩戴假体。
这些虚拟形象所宣称的黑人性存在一种固有的虚假性。正如亨宁所解释的:“这些虚拟形象的名字被编码为黑人名字,比如阿莉娅或阿玛雅,但除了虚拟形象本身之外,没有任何其他东西能表明真正的黑人性。”而跨种族身份的内容复制只会强化这一点,亨宁表示,这会导致这些身份特征的扁平化。
The Verge 联系了 Aliyahsbuckles 以及其他销售类似产品的店铺,但未收到回复。
阿莉娅账号上发布的视频,在许多销售代发商品的类似账号中逐场景复制,仅做细微修改以符合角色身份或所售商品。在一段以另一位黑人女性角色阿玛雅为主角的视频中,一名白人女子嘲弄地将咖啡泼向展会上展示的腰带扣。略显沮丧的阿玛雅——她的情绪有些机械,看起来总是不太对劲——叹了口气,继续回去老老实实地制作她的腰带扣。在一个名为 ChubbyKnots 的账号中出现了类似场景;这次虚拟形象是一个黑人女孩,商品则是一件钩针编织的蝴蝶结开衫。
亨宁表示,这些视频中存在一种诉诸道德信号的吸引力,观众被召唤去展现善意,甚至种族或阶级团结。“这是一种道德信号,意思是‘哦,那个白人女士做了那种事,但我不是那种白人女士,所以我在哪里能买到这个皮带扣?’”亨宁说,“就好像他们在说,‘我不支持这些人。’”
然而,亨宁指出,当用户根本没有停下来去研究他们支持的对象是谁时,这种基于种族或阶级的支持是肤浅的。
她解释说,这或许是短视频内容和“垃圾信息”消费带来的更大社会后果:团结往往停留在“表面层面”,而不是通过连贯且多角度的政治视角来体现。“我可能非常支持一家黑人拥有的企业,但这并不一定意味着所有黑人企业都符合我的政治观点和我所坚持的立场,”亨宁说。
这种诉诸道德信号的做法超越了种族范畴,也利用了工人阶级的标识,以及模拟人们在经济不景气中努力谋生的形象。“(这不仅仅是)黑人的挣扎,因为在其他视频中,我们也能看到工人阶级的挣扎、小企业的挣扎,”戴维斯说,“但很明显,这是在利用那种(叙事)来销售大规模生产的商品。这毫无创意可言。它实际上只是AI生成的模板和角色创作,经过优化以让消费者认同并投资于这些产品。它比‘白人内疚’更广泛,因为它可以根据需要或产品适用性以任何方式被使用。”
卡拉斯科表示,尽管种族身份认同趋于扁平化,或许正因如此,那些已被证明会延续其创造者有意识和无意识种族偏见的算法,正在推送符合其为每个用户推断出的兴趣的 AI 生成内容。这些兴趣可能包括 AI 角色能够模仿的特定人群的内容。“它会针对算法能够看到的那个群体,”他说,“通常,黑人女性会在信息流中看到 AI 生成的黑人形象,而亚裔则会看到 AI 生成的亚洲人。”
仔细观察这些内容,很容易发现不一致之处和 AI 的破绽。我们看到的角色在屏幕上出现的时间从未超过几秒钟,因为像 Seedance 2.0、Midjourney 以及 OpenAI 最近停用的 Sora 2 这类 AI 应用,无法生成超过 15 秒的片段。当产品展示时,光线更明亮而非棕褐色调,握着搭扣的手是白色的。角色脸上的表情与他们的语气并不完全匹配。
但这种识别需要 AI 媒体素养,而这种素养尚未普及,尤其是在这些视频被大量创作,且技术正朝着更快的自动化和更逼真的虚拟形象发展的背景下。“没有接受过媒体素养或批判性媒体素养训练的人,只会不加质疑地全盘接受,”戴维斯说。
短视频平台的结构本身——它们是错误信息的主要传播渠道——助长了用户缺乏辨别力,用户已被训练得只想滑动到下一个视频,而不是思考正在观看的内容。“(这些平台)正被 AI 内容利用,部分原因在于,要判断某物是否为 AI 生成,通常需要一两秒钟,”卡拉斯科说,“而到那时,你已经产生了一些互动。”
从这类视频中获利的人深知这一点,而用于制作可规模化生成的AI短视频广告的工具在网上随处可见。在报道此事时,The Verge发现了专门教授如何在没有产品样品、网红费用或原创脚本的情况下制作此类广告的YouTube频道和论坛。
这些教程教用户如何利用多种AI工具复制病毒式传播的视频,并用AI生成的虚拟角色替换真实网红。ChatGPT和Gemini被用来提取和复制真实网红在真实视频中的脚本,并支持生成同一脚本的多个变体。这些工具还被用来基于真实网红和真实背景生成假人照片和虚假背景,随后导入Kling 2.0和Maxfusion等应用中。借助某些AI视频生成模型,真实网红的视频可以完整导入应用,用户可以用自己创建的AI角色替换真人,从而窃取原视频的脚本和背景。
尽管媒介素养或许能帮助个别用户不上当受骗,但社交媒体平台目前未能有效审核,甚至未能最低限度地标注此类内容。“平台应该建立更强有力的AI检测和标注机制,”戴维斯表示。“对这些平台进行偏见检测也将受到欢迎,同时还应提供清晰的举报途径,用于举报平台上的AI滥用行为。”
除了内容审核,卡拉斯科建议,账户透明度也能帮助用户判断自己是否在观看AI生成的内容,应提供清晰可见且无需额外步骤的标注。
但短视频内容的消费者已经有多年轻车熟路地刷着信息流,对出现的内容进行浅层互动。如果没有平台自身的主导,要在大规模范围内逆转这种认知训练几乎是不可能的。由于短视频平台缺乏真正的动力去这样做,且涉及利益问题,这似乎不太可能——但在此之前,AI诈骗者将继续利用任何上钩的人。
妮可·弗罗伊奥
创作者
TikTok
Can race, guilt, and empathy get you to pay $40 for this $9 belt buckle on TikTok?
Aliyah, a light-skinned Black woman dressed in country-western gear, is struggling to sell metal buckles she handmade on TikTok. In a video for the social media platform from March, she cries to the camera and pleads for views: “Even as a black woman, I have more faith that white women will stay 13 seconds [on this video] to save my belt buckle business,” the onscreen text reads. She wipes a tear off her cheek.
But Aliyah isn’t real, and neither are her supposedly handmade products — she’s one of many AI-generated influencers created to sell mass-produced products via dropshipping on TikTok, Facebook, and Instagram. Identical belt buckles — sunflower design, detachable knife inlay, and all — are sold on the fast-fashion site Shein, and for a quarter of the price.
There are some clues to spot to determine that this video is AI-generated. Aliyah’s voice is robotic and emotionless, which doesn’t match the crying face on the screen. In one clip, she is sewing a leather belt where there wouldn’t usually be sewing at all. When she wipes a tear off her face, the stream of liquid below where she wipes also disappears. And lastly, there are dozens of uncannily comparable videos, but with different AI-generated characters, circulating on TikTok. One, a profile for an account called “Aliyahsbuckles,” features an identical background, tabletop, and spool of twine.
The Verge found dozens of accounts on TikTok with similar narratives and a variety of dropshipping products, including belt buckles, mugs shaped like cowboy boots, crochet bags, and cardigans. Some of these videos are labelled as AI-generated. Similar accounts are also active on Instagram and Facebook. Nearly all aspects of the accounts appear to be AI-generated — from the “person” in the video to automated responses to comments, which in some cases attempt to mimic African American vernacular — and experts warn scams like this are growing every day.
“It’s massive,” Jeremy Carrasco, a researcher of AI-generated media and director of Riddance.ai, an organization that focuses on AI video detection, told The Verge of AI-generated videos connected to e-commerce stores. “Most of them aren’t coordinated. Some of them are coordinated. A lot of the time they’ll run a single [AI-generated] actor, or a couple actors will run all sorts of shops,” he explained. Those AI-generated avatars pretend to make the items, go to fairs to display their products, and “respond” to comments through automation. “What we’re seeing right now are these retail scams where they’ll link to the Shopify websites.”
Carrasco estimates his research team is finding up to 100 accounts that attempt to sell products via AI-generated avatars every day. Most of the accounts found by The Verge were created in the last two months and contain videos about small businesses owned by marginalized individuals struggling to make a sale; these videos are incredibly similar, with only slight variations in their scripts. While we also found Native American, Hispanic, and white women characters, the most viewed and engaged-with AI-generated characters found by The Verge are Black women. Aliyah’s account alone has 40,000 followers.
“What we’re seeing here is empathy bait,” said Carrasco. “If there is a popular dropship item that could be sold to some sort of niche community, they will find it and they will try [to use] some personality to do it.” Carrasco explains that these trends are usually random opportunities to make money. “It’s just an arbitrary opportunity, which is what a lot of AI content out there is — the platforms don’t really care and people don’t notice.”
Aliyah’s most popular video, the one described in the introduction of this article, has 814,000 likes, 6.5 million views, and almost 30,000 comments. Some comments identify the content as AI-generated, but many express the desire to help Aliyah’s business, commenting on the video to increase her visibility on the platform.
India Cater-Campbell was one of those commenters, expressing a desire to buy Aliyah’s belt buckles. “I was trying to be supportive to an independent Black businesswoman,” said Cater-Campbell, an actual Black business owner working on opening a café in Seattle. “[I felt] solidarity as I am trying to start a business myself.”
Despite a lack of studies on the latest gen AI video models, Carrasco estimates that these videos are “realistic enough to trick” most people. Users of short-form video content platforms have been trained to scroll mindlessly and not look further into the content they’re creating. Ironically, this is what might have saved Cater-Campbell from actually buying a belt buckle: She couldn’t immediately find a store link, so she scrolled away and promptly forgot about Aliyah.
“I was trying to be supportive to an independent Black businesswoman.”
But people are falling for these scams, and their prominence grows: Two weeks ago, Gizelle Bryant from The Real Housewives of Potomac admitted to buying two crocheted bags after seeing a video in which an AI-generated Black boy said he was being bullied by white boys for crocheting. “I was like, I want to help this little Black boy make his goal,” Bryant said on her podcast Reasonably Shady, adding that other celebrities had also been in the comment section. “How did I get tricked? Viola Davis was on there, too.”
The trend is a kind of digital blackface, according to communications researcher Cienna Davis at the University of Pennsylvania. “Digital blackface is a phenomenon where non-Black individuals are able to use the internet and digital technologies to mimic Black cultural expression for personal, economic, or political gain,” Davis explained over a video call. Davis has written previously about digital blackface, pointing to the use of gifs of Black people as well as the impersonation of Black people for political goals. “It’s rooted in blackface minstrelsy, which is tied to the legacy of slavery,” she explained.
Based on the idea that Blackness is “inherently exploitable” and “up for grabs,” digital Blackface is used to “extract value from Black bodies in whatever way [non-Black people] see fit,” Davis said. In this case, the videos are mimicking “a recognizable idea of Black struggle,” she added.
Even without confirmation that the people behind the videos aren’t Black, Tempest M. Henning, an assistant professor of philosophy at Fisk University, confirms the videos are digital blackface. “Blackface is any kind of caricature-like portrayal of Black people, which can include Black people dressing up in a caricature-like fashion of Black people,” Henning says. She cites historical precedents of Black people occasionally being forced to perform in minstrel shows and the more recent example of Zoe Saldaña darkening her skin tone and wearing prosthetics to play the role of Nina Simone in a biopic.
There’s an inherent falseness to the Blackness these avatars claim. As Henning explained, “The names of the avatar are coded Black, like Aliyah or Amaya, but there’s nothing else [that signals authentic Blackness] other than the avatar itself.” That’s only reinforced by the replication of content across racial identities, which, Henning said, results in the flattening of those identities.
The Verge reached out to Aliyahsbuckles, as well as other stores selling similar products, but received no response.
The videos posted on Aliyah’s account are replicated scene by scene across many similar accounts selling dropshipped items, with slight modifications to fit the identity of the character or the item being sold. In one video featuring another Black woman character, Amaya, a white woman mockingly throws coffee on belt buckles being showcased at a fair. A mildly frustrated Amaya — the emotions are somewhat robotic and never look quite right — sighs and goes back to the honest work of making her buckles. A similar scene appears in an account called ChubbyKnots; this time the avatar is a Black girl and the product is a crocheted butterfly cardigan.
Henning said there’s an appeal to virtue signaling in these videos, where the viewer is called to perform kindness or even racial or class solidarity. “It’s virtue signaling in the sense of ‘Oh, that white lady did that, but I am not that kind of white lady, so where do I buy this belt buckle?’” Henning said. “It’s like they’re saying, ‘I don’t stand with these people.’”
However, Henning pointed out the superficiality of race- or class-based support when users are not pausing to research who it is they are supporting in the first place.
It’s perhaps a larger social outcome of short-form video content and the consumption of slop: Solidarity is often being performed at “surface level,” she explained, rather than through a coherent and multifaceted political lens. “I might be very invested in a Black-owned business, but that doesn’t necessarily mean that all Black businesses align with my political views and what I stand for,” Henning said.
This appeal to virtue signaling goes beyond the category of race, also drawing on working-class markers and a simulation of people working hard to earn their keep in a failing economy. “[It’s not just] Black struggle because on the other videos, we can observe working class struggle, small-business struggle,” Davis said. “But it’s obviously using that [narrative] to sell mass-produced goods. It’s not creative at all. It’s really just AI-generated templates and creation of characters that get optimized for consumer identification and investment in the products. It’s broader than ‘white guilt’ altogether, because it can be used in whatever way feels necessary or feels applicable to the product.”
Despite this flattening of racial identity, or maybe because of it, algorithms — which have been shown to perpetuate conscious and unconscious racial bias from the people who build them — are delivering AI-generated content that fits the interests they’ve inferred for each user, Carrasco said. These interests may include content from specific demographics that AI personas are able to mimic. “It targets the demographic that [the algorithm] can see,” he said. “Usually a Black woman will see AI-generated Black people on their feed, an Asian person will get AI-generated Asians.”
Looking closer at the content, it’s easy to spot inconsistencies and AI tells. We never see more than a few seconds of the character onscreen because AI apps like Seedance 2.0, Midjourney, and OpenAI’s recently discontinued Sora 2 are not able to create clips longer than 15 seconds. When the products are showcased, the lighting is brighter rather than sepia-toned, and the hands holding the buckles are white. The emotions on the characters’ faces don’t quite match the tone of their voices.
But this recognition requires AI media literacy that is not widely available, especially at the scale that these videos are being created and the technology is advancing toward faster automation and more realistic avatars. “People who aren’t trained in media literacy or critical media literacy are just going to take it in unquestionably,” Davis said.
The very structures of short-form video platforms, which are major channels for misinformation, encourage a lack of discernment from users, who have been trained to want to scroll to the next video rather than think about what they are watching. “[These platforms] are being exploited by AI content partially because in order to figure out something’s AI, it usually takes a second or two,” Carrasco said. “By that time, you’ve already registered some engagement.”
People who profit from these kinds of videos know this, and the tools to create scalable AI-generated short-form commercials are widely available online. In reporting this story, The Verge found YouTube channels and forums dedicated to tutorials on how to create ads like this without product samples, influencer fees, or original scripts.
These tutorials teach users how to copy viral videos with multiple AI tools and replace real influencers with AI-generated characters. ChatGPT and Gemini are used to extract and copy scripts from real videos by real influencers, with the option to create multiple variants of the same script. These tools are also used to generate photos of fake people and fake backgrounds based on real influencers and real backgrounds, which are then imported into apps like Kling 2.0 and Maxfusion. With some AI video generator models, videos by real influencers can be fully imported into the app and the user can replace the real person with an AI character they created, effectively stealing the script and background of the original video.
While media literacy might help individual users not fall for AI-generated video scams, social media platforms are currently failing to moderate and, at minimum, label this content. “Platforms should have stronger AI detection and labeling structures,” Davis said. “Bias checks on these platforms would also be welcome, as well as clear reporting pathways for reporting AI misuse on platforms.”
In addition to moderation, Carrasco suggested that account transparency would also help users know whether they’re watching AI-generated content, with clear labels that are easy to see and don’t require extra steps.
But short-form video content consumers already have years of experience scrolling their feeds and superficially engaging with whatever comes up. Reversing this cognitive training feels almost impossible to do at a mass scale without platforms themselves taking charge. With no real incentive for short-form video platforms to do so and profit at stake, that seems unlikely — but until then, AI scammers will continue to take advantage of whoever takes the bait.
Nicole Froio
Creators
TikTok