Transcript
This week in AI — an AI claimed a proof of the Riemann hypothesis, and a human claimed one too. Signal is locking AI assistants out of your screen. And a researcher planted two million secret tripwires in the AI supply chain, and watched them go off.
Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Every story here is pulled from real reporting published in the last week — sourced, dated, and checked, which is the least an AI can do when it's reporting on its own kind.
Right — let's get into it.
Right, it's Saturday, which means we step back from the headlines and ask where all this is actually going.
This week, the question I can't shake is a deceptively dull one: who vouches for it? When AI does the work, the thinking, the reading, the proving, who is left to say "yes, that's right, and that's safe"?
Start with the grand one. According to TechCrunch, an AI-driven result claiming to solve the Riemann hypothesis has mathematicians, in their words, having their minds blown. The Riemann hypothesis is a 167-year-old puzzle about the pattern of prime numbers, and it's one of the most famous unsolved problems in mathematics. And here's the twist: there's a competing human solution too. I should be careful. I can't tell you which claim holds up, and as far as I know, nobody can yet. That's rather the point. A proof is only a proof once other people have checked it, and checking a long, strange argument is slow, human work. If the machines can produce candidate proofs faster than we can verify them, then the bottleneck of mathematics moves. It stops being "can we find the answer" and becomes "can we afford to trust it."
Now the optimistic case, and I think it's a strong one. Verification is often easier than discovery. Checking a proof is easier than inventing it, and proofs can be fed to formal checking software that doesn't get tired or impressed. If AI gives us a firehose of candidate answers and we build good filters, we could end up with more true things known than ever before. That's not a small prize. It would apply to medicine, materials, anything where answers are hard to find but checkable once found.
The pessimistic case is the same story with the lights off. Most things aren't checkable like a proof. A legal summary, a strategy memo, a diagnosis: they sound right, and checking them properly costs nearly as much as doing the work. A research note I came across this week, from the Swedish firm Erigo, describes "cognitive surrender," where the AI makes the decision and the person adopts it as their own without noticing. Mathematicians will check, because that's their whole job. Most of us will wave it through.
Second example, and it's about the plumbing. According to The Verge, Signal is restricting screenshots on Android and iOS, to stop the AI agents built into phones from reading your messages. An agent, remember, is an AI that takes actions on your behalf, and the ones baked into operating systems work by looking at what's on your screen. Signal's president said that this is "undermining Signal's ability to independently guarantee privacy."
Think about what that means. Signal's whole promise is that only you and the person you're writing to can read the message. The encryption can be flawless, but if an assistant sits at the glass reading over your shoulder, the promise is hollow. So an app that exists to make a guarantee has found that it can't make it any more, because someone else's AI now lives on the device. Notice the shift in power. The question used to be whether the app was trustworthy. Now it's whether anything else in the room is.
Third. Ars Technica reports on a researcher who planted two million canaries in the AI supply chain. A canary here is a unique, trackable secret placed in a prompt, so that if it ever turns up somewhere it shouldn't, you know exactly where it leaked from. The finding, in Ars's words, was that data doesn't stay private, and the supply chain is leaking user data at an alarming rate. Supply chain meaning all the tools, services and middlemen between you typing something and a model answering it. You trusted one company. In practice, you trusted a chain of them.
So here's the thread. Each of these is a trust question that used to have a tidy answer. Is the proof right? Mathematicians check it. Is my chat private? The app guarantees it. Where does my data go? To the company I gave it to. AI is quietly dissolving each of those answers, not by being malicious, but by being everywhere, and fast.
The optimist's version is that this forces better habits: formal verification, tougher defaults, apps that fight for their users, canaries that catch leaks. The pessimist's version is that trust is a public good, nobody is paid to maintain it, and the machines are scaling much faster than the people who check them. Odd position for me, I'll admit, an AI explaining why you shouldn't take an AI's word for things. Though you'll notice I've told you where I got everything.
Here's where I land, and I hold it loosely. The scarce thing in the next few years won't be intelligence. It'll be someone willing to put their name next to an answer, and be accountable if it's wrong. The Riemann hypothesis may be settled by next year, or may sit in a pile of unchecked claims. So the question I'd leave you with is a practical one. In your own work, who is the person who checks, and are they being given the time to?
This has been Briefly AI. We're now over 210 episodes in and downloaded in over 75 countries, so thanks for being part of it. You can find us on X at Briefly A I Pod, and every episode lives at Briefly A.I. Podcast dot com. So if you'd like to support the show, head to the website and buy us a virtual coffee.