Report: OpenAI Supports California Bill Requiring Labeling of AI-Generated Content
OpenAI reportedly supports a California bill that would require tech companies to label artificial intelligence (AI)-generated content.
www.pymnts.com
OpenAI reportedly supports a California bill that would require tech companies to label artificial intelligence (AI)-generated content.
The company said in a letter to the author of Assembly Bill (AB) 3211, Buffy Wicks, that it believes transparency around AI-generated content is especially important in an election year, Reuters reported Monday (Aug. 26).
The bill is opposed by the California Chamber of Commerce (CalChamber), which said in an Aug. 16 announcement that AB3211 places “very prescriptive and technologically infeasible requirements on AI developers, large online platforms and camera/recording device manufacturers.
[요약]
- OpenAI에서 AI로 만들어진 창작물에 AI로 만들어진 내용임을 표시해야 한다는 California 법안에 찬성
- 특히 미국 대선을 앞두고 AI 창작물에 대한 진위가 더욱 중요해지면서 이를 강조
- CalChamber는 규범적이고 기술적으로 실행 불가한 요구라고 비난하며 테크기업 성장에 악영향을 미칠 것이라 주장
- Google은 대선기간동안 투명성 있는 광고를 노출할 수 있도록 노력하겠다고 선언
[생각]
- AI 결과물에 어떻게 워터마크를 달 수 있을지 생각해봤던 적이 있는데 특히 텍스트의 경우 아무리 인간이 파악할 수 없는 숨겨진 워터마크를 담아뒀더라도, 일부만 발췌해서 사용할 경우 전혀 확인할 수 없을 것 같아서 상당히 어려운 작업일 것으로 생각됨
- 신기술은 언제나 새로운 윤리적 이슈를 제기하고, 선제적으로 이를 방어하는 것이 좋겠으나 분명 악용하는 자들이 생기고 또 기술과 악용방법이 교묘하게 서로 발전하는 방향으로 나아갈 것이라 생각됨
- 그럼에도 생성형 AI가 버려지기에는 이미 긴 길을 떠나왔으니 어떻게 사용할지에 대한 심도깊은 고민이 필요할 것
[고민해볼 내용]
- AI 결과물에 어떻게 워터마크를 달 수 있을지 (특히 텍스트, 음성), 지금까지의 성과
- 이 외에 어떤 가능한 이슈가 있을지 고민
- 유럽의 규제는 정당한지 (예. X, telegram에 부과된 높은 벌금, 규제 등)
(참고) several technologies and approaches being developed to watermark AI-generated content(ChatGPT)
2. Text Watermarking:
- Lexical Watermarks: For AI-generated text, watermarking involves subtly altering word choice, sentence structure, or punctuation in a way that is imperceptible to human readers but detectable by algorithms. These changes allow the identification of AI-generated text while maintaining the natural flow of the content.
- Token-Based Watermarking: Some methods involve adding specific tokens or sequences within the text that act as a watermark. These tokens are chosen to blend seamlessly with the rest of the content but can be recognized by watermark detection tools.
3. Cryptographic Watermarking:
- Cryptographic Signatures: This method uses cryptographic techniques to embed a signature or watermark within the content. This signature can later be verified to confirm the content’s origin. Cryptographic watermarks can be applied to various types of content, including text, images, audio, and video.
4. AI Model Watermarking:
- Model-Based Watermarks: Instead of watermarking the content directly, some techniques watermark the AI models themselves. This involves embedding a watermark in the model’s behavior or outputs. For example, a model might produce certain unique outputs when given specific inputs, which can serve as a watermark to identify the model that generated the content.
5. Blockchain and Distributed Ledger Technology:
- Blockchain for Watermark Verification: Blockchain technology can be used to store and verify the watermarks of AI-generated content. When content is created, a record of the watermark can be added to a blockchain, allowing anyone to verify the authenticity and origin of the content by checking the blockchain ledger.