By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Deepfakes Are Surging, But Detection Remains Cheaper and Feasible
- Vijay Balasubramaniyan says deepfakes have surged, with a 1400% increase in the first six months of this year versus all of last year, as voice-cloning tools grew from 120 to 350 and cloning needs only seconds of audio.
- He describes real cases such as the January New Hampshire Biden robocall, bank customers facing a deepfake per day, and news organizations finding 90% of some conflict footage fake.
- He argues deepfakes are detectable, citing 99% detection with a 1% false positive rate, because generators leave artifacts and must balance sounding human against evading detectors.
- He says detection costs about one-hundredth of generation, and that watermarking is limited because attackers won't comply and a watermark was largely stripped in the Biden call.
- He recommends policy that is hard on threat actors and flexible for creators, modeled on CAN-SPAM and Know Your Customer rules, with platforms held accountable for labeling real versus AI content.
IDEAS WORTH REMEMBERING
5 ideasGenerative AI changed the scale and ease of deepfakes, not the basic scam
Scams like the Japanese 'help me grandma' fraud predate AI, but cloning now needs about 3 to 15 seconds of audio, and fraudsters pair voice tools with LLMs that can invent convincing stories.
Deepfake volume is rising sharply in commerce
Vijay says banks went from about one deepfake a month per customer in 2023 to one per customer per day this year, with some big banks seeing one every three hours.
Deepfakes are detectable because AI can inspect far more detail than human ears
Audio has thousands of samples per second, and generators make spectral and time-domain mistakes that a model can catch even when the voice sounds right to people.
Attackers face conflicting objectives
A generator must sound like the target and also evade detection; adding noise to hide artifacts makes speech unintelligible, as in the LeBron James fake, so the goals diverge.
Detection is much cheaper than generation
Vijay says his detection model is about one-hundredth the cost of generation, so attackers would need roughly two orders of magnitude more resources, and they would have to re-architect whole pipelines to evade it.
WORDS WORTH SAVING
5 quotesThere has been a 1400% increase in the amount of deepfakes we've seen this year in the first six months compared to all of last year.
— Vijay Balasubramaniyan
Like, in LLMs, hallucination is a problem, so the fact that you're making shit up is a bad idea. But if you have to make shit up to convince someone- The LLM is created perfect.
— Vijay Balasubramaniyan
Your human ear Can't look at anomalies 8,000 times a second. If it did, you'd go mad, right?
— Vijay Balasubramaniyan
I think the, the simple one-liner answer is they should make it really difficult for threat actors and really flexible for creators, right?
— Vijay Balasubramaniyan
So to detect it is two orders of magnitude cheaper than creation.
— Martin Casado
High quality AI-generated summary created from speaker-labeled transcript.
