Last week at the MIT EmTech Digital conference in London, I participated in a panel focused on how business, government, and academia can collaborate to maximize opportunities and manage challenges associated with advanced AI products.
Alongside ElevenLabs, the panel included leaders from the Alan Turing Institute, the Ada Lovelace Institute, and BT, with MIT Technology Review’s Melissa Heikkilä moderating the discussion.
AI 安全的三管齐下方法
在 ElevenLabs，我们开发音频 AI 技术时始终关注其影响。作为负责 AI 安全的人员，我致力于帮助创作者、企业和用户，同时防止滥用并遏制不法行为者。在此次小组讨论中，我介绍了为让 ElevenLabs 成为更安全、更具创新性的空间所采取的措施，并倡导采用优先应对 AI 安全挑战的策略。这些策略包括：
Provenance: involves distinguishing AI-generated content from real content by understanding its origins. Upstream AI detection tools, such as classifiers, are probabilistic models trained to recognize AI-generated outputs. At ElevenLabs, we’ve developed the AI Speech Classifier that lets anyone upload samples to check if they originate from our platform. We’re also collaborating with Loccus to enhance AI content classification capabilities. Classifiers, however, are not a panacea solution for provenance; they have their limitations. To address them, downstream AI detection methods have emerged, including metadata, watermarks, and fingerprinting solutions. We endorse industry-wide efforts such as cryptographically signed metadata standard C2PA, which present the benefit of being open and interoperable and could enable labeling of AI-generated content across main distribution channels like Instagram or Facebook.
可追溯性：确保 AI 生成内容可追溯至具体用户。在 ElevenLabs，我们的系统可将平台生成的内容关联到来源账户；语音克隆工具仅面向已通过银行信息验证账户的用户开放。重视可追溯性，可确保所有 AI 平台用户为自己的行为负责，并在必要时由执法部门识别身份。
Moderation: which involves defining clear policies on acceptable content and use, and preventing users from generating content that does not comply with such policies. At ElevenLabs, we use automated systems to scan, flag, and block inappropriate content. Human moderators review flagged content to ensure consistent policy enforcement. We are continually advancing our moderation technology to prevent the generation of content that could harm public trust or safety. Open source moderation endpoints, such as the one provided by OpenAI, enable easy integration of prompt moderation into any AI applications.
Working together towards a common goal
While we prioritize safe AI development, no company can tackle AI misuse alone. At the panel, we recognized the need for widespread collaboration among tech companies, governments, civil rights groups, and the public to confront social and environmental changes and ensure a secure digital future. At ElevenLabs, we’re dedicated to maintaining open dialogue in our community and supporting initiatives like the Content Authenticity Initiative and the Elections Safety Accord to combat misinformation.
让 AI 安全面向未来
在此次小组讨论中，我们一致认为，AI 产品必须以安全的方式开发和使用。同时，也应允许其用于富有创意且出乎意料的场景。在 ElevenLabs，我们经常看到平台被用于提升数字内容的无障碍体验，帮助需要音频转录的人士，也帮助因 ALS 和其他健康状况失去声音的人重获声音。要让 AI 应用蓬勃发展，必须提升人们对 AI 媒体的认知，鼓励审慎看待数字内容，推广验证真实性的工具，并向公众和机构普及合乎伦理的 AI 使用方式。
