Steven Chia Investigates Deepfakes During Elections | Talking Point | Full Episode

Key Concepts

Deepfakes, cheapfakes, AI manipulation, disinformation, election integrity, Eleona bill (Elections Integrity of Online Advertising Amendment Bill), POFMA (Protection from Online Falsehoods and Manipulation Act), GANs (Generative Adversarial Networks), C2PA (Content Authenticity Initiative), AI safety, voice cloning, digital avatars, personalized messaging.

Deepfakes and the Singapore Elections

• Introduction of Eleona Bill: Singapore has introduced the Eleona bill to ban deepfakes during elections. This law prohibits digitally manipulated videos and audio of political candidates saying something false.
• Why Now? Deepfake technology has advanced significantly since the early 2000s, making it harder to distinguish between real and fake content.
• Scope of the Law: The law targets realistic-looking content that could deceive people into believing it's true. It does not apply to cartoons, memes, or satire.
• Relationship to POFMA: While Singapore already has POFMA to address falsehoods, the Eleona bill provides a faster mechanism to deal with deepfakes during the short election period (typically 2-2.5 weeks). POFMA processes typically take longer.
• Decision-Making Process: If a candidate believes a video is a deepfake, they can report it. The returning officer then decides whether it needs to be taken down.
• Examples of Content Covered:
  • Realistic scam videos of a former PM asking people to invest in cryptocurrencies.
  • AI-generated live translation that uses a candidate's mannerisms and voice to deliver speeches in other languages.
• Examples of Content NOT Covered:
  • Generated images of Taylor Swift and her supporters saying that they supported Trump.
• GANs and Deepfake Technology: A breakthrough in deepfake technology occurred in 2014 with the invention of GANs (Generative Adversarial Networks), which can mimic human behavior and speech patterns.

Deepfakes in the US and AI Detection

• US Election Concerns: The 2024 US presidential election saw widespread concerns about AI-generated fake content and its potential to manipulate voters.
• State Laws: Around 20 states enacted laws on AI deepfakes in elections.
• Reality Defender: Reality Defender is a cybersecurity firm that helps companies detect deepfakes. Their software analyzes media (audio, video, images, text) and uses color-coded bounding boxes to indicate the likelihood of manipulation. Red squares over faces indicate a high probability of being a deepfake.
• Challenges in Detection: Even experts struggle to identify advanced deepfakes with the naked eye.
• Future of Detection: Future detection methods will focus on content, context, truthfulness, and emotional biomarkers.

Voice Cloning and 11 Labs

• Joe Biden Robocall: An audio deepfake of former US President Joe Biden told voters to skip the New Hampshire primary election.
• 11 Labs: 11 Labs is a company that offers voice cloning software.
• Voice Cloning Process: The software can create a clone of a voice with only 10 seconds of audio.
• Safeguards: 11 Labs has implemented safeguards to prevent abuse, including:
  • Screening users and blocking sign-ups from bad actors.
  • Maintaining a list of celebrities and public figures whose voices cannot be cloned.
  • Requiring voice verification for high-risk accounts.
• Abuse Examples: The software has been used to create fake messages from candidates telling supporters not to vote.

Legislation and Regulation

• Alex Boris's Proposals: New York politician Alex Boris has proposed AI-related bills to combat deepfakes, including:
  • Requiring labeling of campaign materials manipulated with AI.
  • Allowing candidates to sue in court if a deepfake is created of them near an election.
• C2PA Standard: The C2PA (Content Authenticity Initiative) is a file format that stores information about how audio, video, or images were created. This metadata cannot be faked, providing a way to verify the authenticity of content.
• Adoption of C2PA: Most AI image generators attach C2PA information. However, no phones currently have this built-in.

Beneficial Uses of Deepfakes

• Eric Adams's Deepfakes: New York Mayor Eric Adams created audio deepfakes of himself speaking several foreign languages to reach more people.
• BeHuman: BeHuman is a digital cloning company that offers AI avatars for campaigns, enabling personalized video messages to voters.
• Personalized Messaging: BeHuman's technology allows for the customization of video messages, including names and specific campaign issues.
• Limitations: The technology cannot create videos entirely from scratch. It requires pre-recorded video with pauses around the issues to be changed.

Deepfake Experiment and Public Perception

• Experiment Setup: The presenter created a deepfake video of Singaporean Minister Josephine Tio saying the word "otters" (which she never said in the original clip). This video, along with three real videos, was shown to the public.
• Results: People correctly identified the deepfake video, but their reasons were based on visual cues (odd complexion, funny background, blurry image) rather than the manipulated audio.
• Minister Tio's Reaction: Minister Tio expressed concern about the potential for deepfakes to misrepresent candidates and influence election outcomes. She stated that she would not use deepfakes to communicate with residents in different languages, as it would be a misrepresentation.

Cheapfakes

• Definition: Cheapfakes are videos or images manipulated using non-AI technologies (e.g., cropping, slowing down, adding fake captions).
• KTOS: KTOS is a government initiative in Singapore that combats misinformation online and tracks the development of cheapfakes.
• Examples:
  • Taking a photo of refugees out of context and using it to depict a current war.
  • Adding fake sound effects to a video.
  • Cropping images and inserting people into different scenes.
• Concerns: Both cheapfakes and deepfakes are a concern, as they can both be used to deceive people. Cheapfakes are becoming more sophisticated with the integration of AI technologies into editing tools.

Conclusion

• Deepfake Reveal: The presenter revealed that the opening video of the episode was a deepfake.
• Tips for Spotting Deepfakes:
  • Look at the eyes (too little or too much blinking).
  • Check if the lips are syncing with the words.
  • Look for anything unnatural about the face.
• Final Thought: It's important to be aware of deepfakes and cheapfakes and to be critical of the information we consume online.