The Truth About Deep Fakes, and What We Should Do About Them | Ben Colman | TEDxMidAtlantic

Key Concepts:

• Democratization of AI
• Deep fakes
• Generative AI
• Fraud (identity, financial, etc.)
• Provenance vs. Inference
• AI detection of AI
• Regulatory landscape (or lack thereof)
• Perimeter strategy
• Bipartisan support for AI regulation

1. Democratization of AI and its Impact

• Over the last few years there has been a "massive democratization of AI," driven by decreased hardware and software costs. Inference costs have reduced by a factor of 100.
• Over 10,000 AI products are readily available.
• Anyone can create deep fakes, regardless of AI experience, leading to entertaining content but also "nefarious" applications like non-consensual pornography and identity fraud.
• Adoption rates for generative AI are much faster than previous emerging technologies, taking days instead of months or years.

2. Deep Fakes: Statistics and Real-World Examples

• A "1,700% increase in deep fakes" has been observed in the US.
• Less as Nexus CEO forecasts over "$1 trillion of fraud" over the next 12 months.
• Over one-third of large companies have already faced deep fake fraud.
• The use of generative media in fraud has increased "10-fold since last year."
• Examples include fake audio of a president discouraging voting, false claims from doctors about vaccinations, deep fake non-consensual pornography (e.g., Taylor Swift), and high school students using "Nutify applications."
• Deep fakes have been used to influence elections outside the US, such as a deep fake audio of a candidate claiming that he was going to increase the taxes on beer.

3. Challenges in Identifying Deep Fakes

• Traditional methods of deep fake detection (symmetry analysis, anti-aliasing, sound-mouth correlation) are no longer effective.
• Even experts (e.g., two-thirds of Reality Defender's team) can't reliably distinguish deep fakes.
• Current platforms focus on "provenance," or matching like for like using methods like watermarks, fingerprints, and community notes.
• Community notes are often too slow, appearing only after content has gone viral (e.g., fake Pentagon explosion causing a "$100 billion flash crash").

4. The Inference Approach to Detection

• Reality Defender focuses on "inference" rather than provenance.
• Inference is "probabilistic" and assumes the "ground truth may not ever exist."
• Uses a "perimeter strategy," considering multiple factors to determine if media has been AI-generated or manipulated.

5. Regulatory Landscape and Advocacy

• There are "no current federal regulations on generative AI fraud," with only limited state and city-level regulations that reiterate existing laws.
• The speaker advocates for regulations that require service providers to protect consumers (akin to banks requiring metal detectors).
• Reality Defender drafted its own regulatory proposal that supports AI innovation while requiring improper use of AI to be clearly marked.
• Testimony was given before the Senate Judiciary Committee, where bipartisan support for AI regulation was evident.

6. The Imperative of AI Regulation

• In a year where two-thirds of the world is voting in elections, regulations are needed to protect the integrity of these elections.

7. Notable Quotes/Statements

• "Massive democratization of AI"
• "1,700% increase in deep fakes"
• "Two-thirds of my team...can't tell the difference."
• "We must use AI to detect AI."
• "Your voice is your password."

8. Technical Terms/Concepts

• Democratization of AI: The increasing accessibility and affordability of AI tools and technologies.
• Deep fakes: AI-generated media (images, videos, audio) that convincingly impersonate real people or events.
• Generative AI: A type of AI that can generate new content, such as text, images, or audio.
• Provenance: Verifying the origin and authenticity of a piece of media.
• Inference: A probabilistic approach to deep fake detection that considers multiple factors.
• Perimeter strategy: A comprehensive approach to security that involves multiple layers of defense.

9. Conclusion

The democratization of AI has led to a surge in deep fakes and AI-generated fraud. Traditional methods of detection are failing, necessitating AI-powered inference-based solutions. The lack of federal regulations poses a significant threat, particularly to upcoming elections, and requires immediate attention and bipartisan action to protect consumers and ensure the integrity of information. The focus should shift from basic checks to inference.