Content Authenticity Initiative reposted this
Think you're good at spotting AI-generated voices? Try the quiz below, which suggests your accuracy is likely not much better than a coin toss. A study by Sarah Barrington and Hany Farid tested 50 participants who each listened to 40 short voice recordings, half real, half AI-generated. The average accuracy was only 65%. This initial sample size is small and the voice clips that participants heard were short (between 3 and 10 seconds), but it nonetheless supports my position that emphasising human detection of synthetic media is the wrong approach. Advances in AI-generated voice have been dizzying in the last two years, arguably the most dramatic across all modes of synthetic media. Just a few years ago, Google's TacoTron 2 was the leading accessible voice cloning tool but was far from generating highly convincing outputs. Today, there's no shortage of startups and open source projects pushing forward the frontier, realistically synthesising the sound of a person's voice, but also how they speak naturally and fluidly. There are still subtleties and limitations, but don't forget these are still early days. Media literacy is often brought up in response to the challenges of deepfakes and synthetic media. Awareness is important, but we can't (in good faith) say "Here's how you can spot AI-generated content reliably" without giving false confidence over time. There may be some tells today, but the direction of travel is clear. Voice synthesis is, as Hany says, "passing through the uncanny valley". We need to recentre our discussion of 'detecting deepfakes' and responsibility for doing so away from the individual. Automated detection & AI classifiers have a role to play, but media provenance approaches such as Content Authenticity Initiative and Coalition for Content Provenance and Authenticity (C2PA)'s Content Credentials enable trust and transparency to be 'baked in' from the bottom-up. Voice synthesis leaders such as ElevenLabs, Microsoft, OpenAI, and Respeecher have all joined CAI which is promising. The next step is widespread adoption. https://lnkd.in/e_D-cBRY #ai #deepfakes #generativeAI #syntheticmedia