内容来源：https://www.wired.com/story/europeans-are-about-to-find-out-how-entrenched-ai-is-in-their-daily-lives/

内容总结：

欧盟《人工智能法案》8月2日正式生效，要求企业在用户与AI系统互动或接触AI生成内容时明确告知。新规旨在减少欺骗和操纵，增强透明度，涵盖广告、聊天机器人、客服热线、日程管理等多个商业场景。违规企业将面临最高1500万欧元或全球年营业额3%的罚款。

法案同时适用于AI模型开发者，由欧盟AI办公室监督，涉及OpenAI、Anthropic、Google DeepMind等公司，也包括Spotify、Adobe等使用AI功能的企业。批评者担忧过度标注将导致信息疲劳，使标注失去意义。该法案与2018年GDPR类似，面临执行标准不清晰、技术实施复杂等挑战。欧盟已给予过渡期至12月，但各国监管框架进度不一，执法可能不会立即统一。律师认为，新规将迫使企业重新审视AI使用策略，可能成为行业分水岭。

中文翻译：

广告、音乐推荐、日历管理应用以及投诉热线——如果它们使用人工智能，欧洲人将需要被告知这一情况。

从本周日开始，欧盟公民在与人工智能系统互动或查看由该技术生成或修改的内容时，必须被告知。这些新的透明度规定是欧盟《人工智能法案》的一部分，该法案将于8月2日生效。该法案旨在减少欺骗和操纵，帮助人们做出明智的选择，其更广泛的目标是在欧洲建立值得信赖的人工智能。

“如果提供商真正遵守透明度义务，人们将非常清楚地看到人工智能被使用了多少，”Stibbe律师事务所的技术律师弗雷德里克·费尔诺特告诉《连线》杂志。“尤其是在营销领域。”

使用深度伪造技术的广告或企业社交媒体帖子将被贴上标签，注明其为人工智能生成。处理投诉和其他问题的聊天机器人和热线必须明确说明它们是否基于人工智能。使用机器学习来监控来电者情绪并检测沮丧情绪的呼叫中心，必须在通话开始时声明其正在这样做。

即使在企业之间，任何对该技术的商业用途，包括安排预约、处理通信或谈判合同，都必须予以声明。如果企业不遵守规定，将面临最高1500万欧元（约合1700万美元）或全球年营业额3%的罚款——以较高者为准。

模型开发者也将受到新法律的约束，并受到欧盟委员会最近设立的人工智能办公室的监督。这包括除了OpenAI、Anthropic和Google DeepMind这些明显例子之外的公司。Stibbe律师事务所的技术和数据律师蒂博·杜坎指出，像Spotify这样使用人工智能推荐的公司，或Adobe在其Photoshop中使用人工智能编辑功能的公司，都属于此列。

“当你开始留意时，你会发现它无处不在，从现在起所有这类内容都必须贴上标签，”杜坎说。

批评者表示，新法律的覆盖面之广可能导致信息过载。“过度标注可能导致‘横幅盲区’，即无休止的通知让人们视而不见，”计算机与通信行业协会的人工智能政策负责人博尼法斯·德·尚普里在12月警告说。“如果我们必须给所有东西贴标签，从简单的拼写检查邮件到带滤镜的照片，那么人工智能内容的标注将失去所有意义。”

这项新立法与欧盟2018年生效的隐私法《通用数据保护条例》（GDPR）有相似之处。该法规赋予消费者了解和掌控公司如何使用其个人信息的新权利，改变了人们体验网络的方式。它引入了无处不在的饼干同意弹窗，让许多网民患上“饼干疲劳”，并导致企业争相合规。

费尔诺特表示，采用和执行此类技术法规的挑战也同样存在。“仍然有很多术语不明确，需要当局提供具体指导，我认为这里也会出现同样的情况，”她说。“它还非常技术性，因此这些要求也需要内置到底层系统中，这使得事情变得更加复杂。”

欧盟已允许一个过渡期，直到12月，让人工智能模型提供商以机器可读的格式标注合成的音频、图像、视频或文本，以便这些材料可以被检测为人工智能生成，这明确承认了实施这一点的技术挑战。该法案中的其他合规期限已被推迟，因为欧洲人工智能办公室在制定关键合规指南方面花费的时间比预期更长。

欧盟的27个成员国在建立各自的监管框架方面也处于不同阶段，这意味着执法“可能不会像我们想象的那样从第一天起就立即或统一，”诺顿罗氏富布赖特律师事务所的数据和人工智能监管律师罗西·南斯告诉《连线》杂志，并警告说“我们不一定马上就能看到一致的执法。”

尽管如此，该法规的广泛覆盖面最终将成为企业的“游戏规则改变者”，杜坎说。“他们会愿意为其贴上标签，还是转而使用非人工智能的系统？他们会回到手动操作，还是愿意冒风险透明地披露他们使用了多少人工智能，继续在所有事情上使用人工智能并如实披露？”

英文来源：

Advertisements, music recommendations, calendar management apps, and complaint hotlines—if they use AI, Europeans will need to be told.
Starting on Sunday, citizens of the European Union must be informed when they are interacting with an AI system or viewing content generated or altered by the technology. These new transparency regulations come as part of the bloc’s Artificial Intelligence Act, which goes into effect August 2. The act is intended to reduce deception and manipulation, and help people make informed choices within the wider aim of fostering trustworthy AI in Europe.
“If a provider’s really compliant with the transparency obligations, it will become very visible how much AI is used,” Frederiek Fernhout, a technology lawyer at Stibbe, tells WIRED. “Especially in marketing.”
Ads or corporate social media posts using deepfakes will be slapped with a label stating they are AI-generated. Chatbots and hotlines handling complaints and other issues will have to state clearly if they are AI-based. Call centers using machine learning to monitor caller emotions and detect frustration will need to declare they are doing so at the start of the call.
Even among businesses, any commercial use of the technology, including scheduling appointments, handling correspondence, or negotiating contracts, will have to be declared. If companies don’t comply, they will face fines of up to €15 million ($17 million) or 3 percent of worldwide annual turnover—whichever is higher.
Model developers will also be subject to the new law, and will be monitored by the European Commission’s recently established AI Office. That includes companies beyond the obvious examples of OpenAI, Anthropic, and Google DeepMind. Thibau Duquin, a technology and data lawyer at Stibbe, pointed to firms like Spotify, which uses AI-powered recommendations, or Adobe for its AI editing features in Photoshop.
“When you start looking, it’s everywhere, and all this content will have to be labeled from now on,” says Duquin.
The reach of the new law could lead to information overload, critics say. “Excessive labeling can cause ‘banner blindness’ with endless notifications,” Boniface de Champris, AI policy lead at the Computer & Communications Industry Association, warned in December. “If we have to label everything, from simple spell-checked emails to photos with a filter, the labeling of AI content will lose all meaning.”
The new legislation is resonant of the EU’s privacy law, the General Data Protection Regulation (GDPR), which came into force in 2018. In granting consumers new rights to understand and control how companies used their personal information, the regulation changed how people experience the web. It introduced ubiquitous cookie-consent pop-ups, leaving many web users with “cookie fatigue,” and caused a scramble by businesses to comply.
The challenges of adopting and enforcing such technical regulations resonate too, says Fernhout. “Still, a lot of terms are not clear and need specific guidance from the authorities, and I think the same will happen here,” she says. “It is also very technical so these requirements need to be built into the underlying systems as well, which complicates it.”
The EU has allowed for a transitional period, until December, for AI model providers to label synthetic audio, image, video, or text in a machine-readable format so that this material can be detected as AI-generated, clearly recognizing the technical challenge of implementing this. Other compliance deadlines within the act have been delayed, as the European AI Office has taken longer than expected to produce key compliance guidelines.
The bloc’s 27 member countries are also at different stages of setting up their supervisory frameworks, meaning enforcement “might not be as day one immediate or uniform as we think," Rosie Nance, a data and AI regulatory lawyer at Norton Rose Fulbright, tells WIRED, cautioning “we’re not necessarily going to see consistent enforcement straight away.”
Nevertheless, the sweeping reach of the regulation will eventually be a “gamechanger” for companies, Duquin says. “Will they be comfortable with labeling it, or will they go back to using systems that are not AI-based? Will they go back to doing things manually, or are they willing to take the risk of being transparent about how much AI they use, and continue to use AI for everything and disclose it as such?”
Comments
Back to top