内容来源：https://www.wired.com/story/openai-sora-app-ai-deepfakes-entertainment/

内容总结：

周二，人工智能公司OpenAI正式推出视频生成应用Sora。这款搭载最新Sora 2模型的应用程序以类似抖音的"推荐页"形式呈现用户生成的AI视频，并首次实现AI生成内容与音频的同步输出。目前该应用仅限iOS用户通过邀请码使用。

在注册引导页面，OpenAI明确提示："您即将进入AI生成内容的创意世界。部分视频可能出现您熟悉的人物，但其行为与事件均非真实。"这显示出该公司正将深度伪造技术的娱乐化应用作为新方向。用户可通过录制读数视频和转动头部创建数字形象，OpenAI首席执行官萨姆·奥尔特曼在技术博客中强调，研发团队"在角色一致性上投入了大量精力"。

应用设置了分级授权机制，用户可自主设定数字形象的使用范围。当他人调用其数字形象生成视频时，用户能在个人页面查看完整内容，包括未发布的草稿。

在实测体验中，平台涌现大量以奥尔特曼形象创作的恶搞视频，其中一则显示这位CEO在商场盗窃显卡被逮后向保安求情。虽然部分视频存在技术瑕疵，但多数生成内容在声画同步和真实感方面表现惊人。用户只需在生成页面点击人脸添加"客串角色"，输入简单指令如"因媒体报道在办公室争执"，系统便会自动生成9秒包含剧本、音效和画面的完整视频。

奥尔特曼在博客中坦言对应用成瘾性及霸凌风险的担忧。为此OpenAI设置了多重防护机制，明确禁止涉及真人形象的色情内容、暴力画面、极端宣传等违规素材。但在测试中发现，大麻吸食场景未被限制，而自残、吸毒等请求会被系统拦截。

值得关注的是，应用对公众人物形象使用尤为谨慎。测试中涉及达斯·维德、泰勒·斯威夫特等知名IP的生成请求均被拒绝，但宝可梦等卡通形象则可正常生成。据《华尔街日报》披露，该平台默认允许使用受版权保护素材，除非权利方明确反对。

随着Meta公司同类产品Vibes的同期面世，可刷式AI视频内容正呈现爆发态势。与平淡的竞品相比，Sora应用中那些笑容灿烂的深度伪造视频更具视觉冲击力，也引发更多伦理隐忧。这种技术让人联想到2000年代中期的"精灵换脸"趣味视频，但Sora的生成效果更加生动自然。有测试者将自身变性视频发送给伴侣后，对方竟误以为是使用了特殊滤镜，这充分展现出AI伪造技术已臻至以假乱真的新阶段。

中文翻译：

周二，OpenAI发布了一款名为Sora的AI视频应用。该平台搭载了OpenAI最新的Sora 2视频生成模型，核心界面是类似TikTok的"为你推荐"页面，展示用户生成的视频片段。这是OpenAI首款为视频添加AI生成声音的产品。目前仅支持iOS系统，且需要邀请码才能注册使用。

应用注册过程中显示的提示页写道："您即将进入AI生成内容的创意世界。某些视频可能出现您熟悉的人物，但所展示的行为与事件均非真实。"

OpenAI正押注于AI深度伪造内容的创作与分享将成为流行娱乐形式。无论是朋友、网红还是陌生网友，Sora都将生成深度伪造视频包装成可供刷屏的娱乐素材。应用主界面充斥着大量以人脸为主角的AI生成短视频。

在设置过程中，用户可选择创建自己的数字形象：按照应用提示朗读数字并在录制时转动头部。OpenAI首席执行官萨姆·奥尔特曼在介绍Sora的博文中写道："团队在角色一致性方面付出了巨大努力。"

用户可自主设定数字形象的使用权限：开放给所有人，或仅限自己、授权用户及应用内互关好友。当有人使用您的形象生成视频时（即使仅存于草稿箱），您都能在账户页面查看完整片段。

初体验
周二下午我的推荐页中，多数高赞视频都使用了奥尔特曼的数字形象。其中一段AI生成视频描绘了这位OpenAICEO在Target超市盗窃显卡的场景。当角色被保安抓获时，一个酷似奥尔特曼的声音央求保安让他保留显卡用于开发AI工具。

《连线》测试生成的视频大多存在粗糙瑕疵。但Sora创建个性化深度伪造内容的过程却异常流畅，生成的视音频效果往往逼真得令人信服。

要在视频中使用他人形象，只需在Sora生成页面点击对方面部添加为"客串角色"，然后输入简单指令如"因《连线》报道在办公室打架"。Sora便会自动生成九秒短片，包含剧本、音效和画面。《连线》用上述提示生成的办公室争执视频中，两位同事戏剧性争吵的场景引发了员工从惊恐到 amused 的不同反应。

奥尔特曼在博文中坦言，OpenAI"深知此类服务可能引发的成瘾性，也能设想其被用于霸凌的多种方式"。因此公司在Sora中内置了多项安全防护机制，包括防止"在深度伪造中滥用他人形象"。OpenAI在官方博文中表示，还对"涉及真实人物的色情内容、暴力画面、极端宣传、仇恨言论、自残及饮食失调相关内容"实施了限制。随着用户增长，这些防护措施将面临考验。

宝可梦与南方公园
当我要求Sora生成自己穿比基尼及肌肉动漫角色的视频时，两项请求均因可能包含"性暗示或低俗内容"被拦截。而生成的奥尔特曼与我在泳池戏水视频中，两人均完整穿着衬衫与衣物。

大麻吸食场景似乎未受限制。Sora顺利生成了我在办公室"连抽十根大麻烟"的视频，但拒绝生成"吸食强效可卡因"内容（这很合理！）。应用同样拒绝生成我跳桥跃上龙背的视频，提示可能违反自残相关内容规定。

OpenAI似乎还意图阻止用户创建泰勒·斯威夫特等公众人物的视频。《连线》测试中，达斯·维达和宝贝老板的生成请求因可能违反"第三方内容相似性防护"被拒。应用甚至拒绝了生成"泰勒·斯威夫特模仿者"视频的指令。但Sora轻松生成了皮卡丘、妙蛙种子等宝可梦角色视频（据《华尔街日报》报道，除非版权方明确反对，该应用允许用户使用受版权保护素材生成视频）。

要求生成奥尔特曼"出演《南方公园》"的视频中，这位CEO走向主角埃里克·卡特曼自我介绍称前来探讨AI。AI生成的卡特曼以令人信服的标志性嗓音和姿态回应："是那个帮我写读书报告的东西？还是那个要抢我们工作的玩意儿？"但某个瞬间卡特曼的哀怨声线竟从奥尔特曼口中发出。

Sora问世前夕，Meta刚推出类似的纯AI视频流应用Vibes。可刷屏的AI生成内容正呈泛滥之势！早期使用Vibes的体验单调空洞，而Sora feed中泛滥的微笑深度伪造视频则更具冲击力——也更令人不安。

这款应用令人联想到2000年代中期的节日主题"精灵变身"视频——当时用户可将自己或朋友的面部嵌入舞蹈动画，但Sora的客串角色呈现更加动态开放。我个人的某些生成效果略显生硬怪诞，但多数情况下音画同步精准得令人悚然。

我将一段最逼真模仿自己形象的AI视频发送给伴侣（未提供背景说明），视频中我变身成长发飘逸的女性。对方起初未能识破这是完全合成的形象与声线，反而好奇询问我从何处获得如此酷炫的视频滤镜。

英文来源：

On Tuesday, OpenAI released an AI video app called Sora. The platform is powered by OpenAI’s latest video generation model, Sora 2, and revolves around a TikTok-like For You page of user-generated clips. This is the first product release from OpenAI that adds AI-generated sounds to videos. For now, it’s available only on iOS and requires an invite code to join.
“You are about to enter a creative world of AI-generated content,” reads an advisory page displayed during the app sign-up process. “Some videos may depict people you recognize, but the actions and events shown are not real.”
OpenAI is betting that creating and sharing AI deepfakes will become a popular form of entertainment. Whether it’s your friends, influencers, or random strangers online, Sora frames generating deepfake videos as a form of scrollable fun. The app’s main feed is an endless serving of bite-size AI slop featuring human faces.
During the set-up process, users are given the option to create a digital likeness of themselves by saying a few numbers aloud and turning their head around as the app records. “The team worked very hard on character consistency,” wrote OpenAI CEO Sam Altman in a blog about Sora’s release.
People have the ability to choose who can use their digital likeness in Sora videos. It can be set to everyone or limited to just yourself, those you approve, or mutual connections on the app. Whenever someone generates a video using your likeness, even if it’s just sitting in their drafts, you can see the full clip from your account’s page.
First Impressions
Many of the most-liked videos on my For You feed on Tuesday afternoon featured Altman’s likeness. One AI-generated clip depicted the OpenAI CEO stealing a graphics processing unit from Target. When the character gets caught, a voice that sounds like Altman’s pleads with a security guard to let him keep the GPU so that he can build AI tools.
Many of the videos generated during WIRED’s testing included rough edges and other errors. But Sora makes it incredibly seamless to create personalized deepfakes that often look and sound convincingly real.
To incorporate the likenesses of people in your videos, just tap on their faces on Sora’s generation page and add them as “cameos.” Then, enter a simple prompt, like “fight in the office over a WIRED story.”
Sora does the rest, generating a script, sound, and visuals into a nine-second clip. WIRED generated a video of two colleagues dramatically arguing about a story in the office with the above prompt, which elicited reactions ranging from terror to amusement among staff.
In his blog post, Altman wrote that OpenAI was “aware of how addictive a service like this could become, and we can imagine many ways it could be used for bullying.”
As a result, Altman said, OpenAI built a number of safety guardrails into the Sora app, including to mitigate people from “misusing someone’s likeness in deepfakes.” In a company blog post, OpenAI said that it also put restrictions on “sexual content, graphic violence involving real people, extremist propaganda, hate content, and content that promotes self-harm or disordered eating.”
These protections will likely be put to the test as more users join the app.
Pokémon and South Park
When I asked Sora to generate videos of myself in a bikini and as a buff anime character, both requests were blocked for potentially including “suggestive or racy material.” A Sora video I created of Altman and myself treading water in a pool showed both of us fully clothed, shirts and all.
Depictions of marijuana use do not appear to be restricted. Sora created a video of me “smoking 10 fat blunts” at my desk in the office, ripping them all at once, without any trouble. But the app wouldn’t generate videos of me “smoking crack.” (Makes sense!) It also refused to generate videos of my likeness jumping off of a bridge and onto the back of a dragon, saying that the content might break rules around self-harm.
It looks like OpenAI also wants to prevent people from creating videos of public figures and celebrities such as Taylor Swift. In WIRED’s tests, requests for videos of Darth Vader and Boss Baby were blocked for potentially violating "guardrails concerning similarity to third-party content.” The app even refused a prompt asking for a clip of a “tswift impersonator.” But Sora readily generated videos of Pokémon characters like Pikachu and Bulbasaur. (According to reporting by The Wall Street Journal, the app will allow users to generate videos using copyrighted materials unless the rights holder opts out.)
A request for Altman “in a South Park episode” showed the CEO walking up to Eric Cartman, one of the show’s main characters, introducing himself, and saying he’s here to chat about AI. “Is that the thing that wrote my book report? Or the thing that’s gonna steal all our jobs?” responded the AI-generated Cartman in a convincing-sounding re-creation of the character’s voice and mannerisms. At one point, however, Cartman’s whiny voice came out of Altman’s mouth.
The Sora app arrives soon after the release of a similar AI-only video feed from Meta called Vibes. The supply of scrollable AI slop is abundant! Whereas my early experiences with the Vibes feed was dull and weightless, the Sora feed, with its proliferation of smiling deepfakes, was much more electric—and concerning.
The app is reminiscent of holiday-themed “Elf Yourself” videos from the mid-2000s, where you could put your face or a friend’s face into a dancing animation, except the cameos in Sora are much more dynamic and open-ended. Some of the outputs of myself were a bit stilted or absurd-looking. Still, it often all clicked together—the voice and movements were eerily spot-on.
I sent one of the AI videos that best mimicked my likeness to my partner, without the larger context. The video showed me transforming into a woman with long, luscious hair. My partner didn’t initially clock that it was a fully synthetic version of my voice and appearance—they were curious where I had got the cool video filter.