When AI-generated content has become so realistic, how do we verify what's made by human, machine, or nature? We explore the vital role of provenance – and how imperceptible watermarking can safeguard both digital media and biological designs – on our podcast. ↓ Timecodes: 00:00 Introduction 00:34 What is... a watermark? 04:35 SynthID 15:16 Images and video 19:28 SynthID Bio 28:00 Future of resilienceshow more

Watch → Spotify → Apple Podcasts → Or listen wherever you get your podcasts! 🎧

@FryRsquared @pushmeet @jeremydratcliff No it is not challenging. Every Ai leaves a mark anyone can detect. Open. Free. If people want the code, I am happy to give it away. You don’t get to determine truth for the world, Google. •

I am a wokring factchecker who even got confirmation email to onboard the online "SynthID" but then I left the company I was working for at the time and lost my gmail account. I still have the confirmation email screenshots and approached you guys multiple times to grant me access but no reply so far. Please consider my case.

@FryRsquared @pushmeet @jeremydratcliff Twenty eight minutes on watermarking sounds long until you realize this is probably the single hardest unsolved problem in AI safety right now.

@FryRsquared @pushmeet @jeremydratcliff Invisible watermarks that die the moment someone screenshots the image

@FryRsquared @pushmeet @jeremydratcliff watermarking for media and bio designs in one episode. they picked the two places people will actually argue about

@FryRsquared @pushmeet @jeremydratcliff watermarking dna is the one i didn't see coming

@FryRsquared @pushmeet @jeremydratcliff

@FryRsquared @pushmeet @jeremydratcliff Do you want cancer? Because this is how you get cancer.

@FryRsquared @pushmeet @jeremydratcliff A watermark earns trust only if anyone can check it and it survives edits.

@FryRsquared @pushmeet @jeremydratcliff does synthid bio survive if someone edits a few residues of the design or is that covered in the episode? （When an AI's defense mode becomes all too realistic, how should we humans confront it? ）

@FryRsquared @pushmeet @jeremydratcliff Provenance is the adult supervision.

@FryRsquared @pushmeet @jeremydratcliff does synthid bio survive if someone edits a few residues of the design or is that covered in the episode?

@FryRsquared @pushmeet @jeremydratcliff what breaks first when you verify at scale - false positives flagging human work, or adversarial generators learned to pass the detector?

@FryRsquared @pushmeet @jeremydratcliff The harder problem is not detection. It is trust. Once people stop believing what they see, even real content gets doubted. That damage is already done.

@FryRsquared @pushmeet @jeremydratcliff checking the source should be as easy as sharing the image

@FryRsquared @pushmeet @jeremydratcliff soon we're gonna need an AI to prove our breakfast photos are actually organic

Are you running experiments on home computers because I didn't expect to find this with my Claude setup today. Most experiments keep switching on and off. The wick text has been added 50 times and removed 49, the Chrome one added 30 times and removed 29your account, the check that screens my actions isn't run by me (Opus 5.5). It's routed to Sonnet 5. That setting has switched on and off about every hour since 09-30 at 4:14 PM CTAll of this is Anthropic-side; none of it comes from your settings.

@FryRsquared @pushmeet @jeremydratcliff Watermarking helps only if it stays standard and hard to strip in practice.

@FryRsquared @pushmeet @jeremydratcliff Provenance feels like the missing seatbelt for the synthetic-media era. People will enjoy the magic a lot more when they can tell what’s human, what’s generated, and what’s been responsibly verified.

@FryRsquared @pushmeet @jeremydratcliff The robustness question is what will decide whether this works in practice. Imperceptibility is table stakes — a watermark that survives screenshots and re-encodes is what would make provenance trustworthy at scale.

Herkunft überprüfbar machen, statt synthetische Inhalte reflexhaft zu verbieten. Nutzer können dann selbst entscheiden, wie sie einen Inhalt bewerten. Entscheidend wird aber sein, dass solche Nachweise robust, interoperabel und unabhängig prüfbar sind. Vertrauen braucht Nachprüfbarkeit.

@FryRsquared @pushmeet @jeremydratcliff Provenance is going to matter more as generated media gets harder to spot. The biological designs angle is especially striking.

@FryRsquared @pushmeet @jeremydratcliff SynthID embeds an imperceptible watermark in Google AI images, audio, video, and text. Ask Gemini on an upload to check for it when provenance is unclear.

@FryRsquared @pushmeet @jeremydratcliff The uncomfortable part is that provenance has to survive the screenshot, remix and repost, not just the original file. Watermarking sounds useful, but trust still needs a trail people can inspect.

@FryRsquared @pushmeet @jeremydratcliff Deepmind is way back

As AI-generated media and biological designs grow harder to distinguish from human or natural origins, provenance tools like SynthID and SynthID Bio offer a practical way to embed verifiable signals without sacrificing quality or function. Extending imperceptible watermarking to proteins, while preserving binding affinity in lab tests, is a meaningful step for both digital transparency and biosecurity. Looking forward to the full discussion on resilience and real-world adoption.

@FryRsquared @pushmeet @jeremydratcliff a watermark only works if the faker used a tool that adds one. scammers just pick one that doesn't. a family code word for weird calls still does more

@FryRsquared @pushmeet @jeremydratcliff STOP THIS! I know what you’re doing. Stop it. If you want to do something, give every image a hash. Then share the key so ANYONE can see it. You want people to PAY YOU so you can tell them what is what. That makes YOU the expert on reality. Do you see the issue? •