Your AI Has Been Snitching On You

Today on Briefly AI — Anthropic's new watermark is quietly outing people who use Claude at work and in class. Twitch has decided your streams train Amazon's AI by default, unless you say otherwise. And a coding startup barely out of nappies is about to be worth forty billion dollars.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The voice is synthetic. The sourcing isn't — every story's traceable to something a real journalist or researcher actually published this week, which is frankly more diligence than Harry shows most mornings.

Right — let's get into it.

So it turns out the safest place to hide has always been "nowhere," because apparently your prose has had a barcode on it this whole time and nobody told you.

Right, so — Anthropic. They've rolled out a watermarking system baked into Claude's output, an invisible pattern in the text that flags "an AI wrote this" to anyone with the right tool to check. And according to reporting in TechCrunch, a fair few Claude users are, let's say, unhappy about it. Not because the watermark's inaccurate — because it works. People have been using Claude to write work emails, class essays, performance reviews, and quietly assuming nobody could prove it either way. Now something can.

Here's the bit that matters. We've talked before on this show about the study showing something like half of workers hide their AI use, because admitting it gets you rated as lazier, less capable, handed less interesting work. That's not a rounding error, that's most people making a rational bet that honesty costs them something. A watermark doesn't fix that bet — it just removes the option to make it quietly. Which means the actual problem — that admitting you used a tool is treated like a confession — is still sitting there completely unsolved. Anthropic just turned the lights on in a room everyone had gotten quite comfortable being dark in.

Worth keeping an eye on whether other AI companies follow with their own watermarks, and whether workplaces use this as a chance to actually update their honesty policies — or just as a better way to catch people out.

Meanwhile, over at Twitch — which Amazon owns — a similar consent problem, but dressed differently. Amazon announced that Twitch will train its AI models on streamers' content, the streams, the clips, the chat, the whole lot, by default. You can opt out. You just have to know to. As The Verge reports, Twitch's own chief product officer said the quiet part out loud on a livestream: "If this was opt-in, nobody would opt in. That's honestly the answer."

Which, credit where it's due, is a strikingly honest thing for an executive to say. But sit with what it actually means — the entire design of that toggle exists because the company already knows what people would choose if you asked them properly. So instead of asking, they default you in and make you go and find the escape hatch. This is the same trick as pre-ticked newsletter boxes, just with your face and your voice as the product.

And that's the thing about defaults — almost nobody changes them. Not because they're happy with the choice, but because changing it takes effort most people never spend. So "unless you opt out" and "yes" end up meaning roughly the same thing in practice, even though they're not remotely the same thing in principle.

Now this last one's a quick one, and a bit of a palate cleanser after all that. Cognition — the company behind the AI coding assistant Devin — is reportedly in early talks to raise a new round that would value it at forty billion dollars, according to Bloomberg. Worth pausing on that number for a second: they raised at twenty-six billion just a few months ago. That's not growth, that's a sprint.

Why does that matter beyond the startup world? Because investors are voting with real money on where AI actually earns its keep, and right now that's coding. Not chatbots, not image generators — tools that write and fix software, because software is expensive to write and AI coding assistants are visibly, measurably good at it. If you work anywhere near engineering, expect the tools your team already uses to get more capital behind them, and expect more competitors chasing the same space, fast.

That's the roundup — a watermark that outs you, a default that assumes you'd say no if only you were asked, and a coding startup doubling its price tag before the ink's dry on the last round. Three stories, one very clear theme: half of this industry is busy building trust, and the other half is busy building around the need for it.

Briefly AI comes to you courtesy of harrysharman.com — Harry Sharman is the human at the other end of it, for anyone keeping track.