One safety story a week, pressure-tested for what's actually happening underneath the surface.

By Lindsay Gross

"If this was opt-in, nobody would opt in. That's honestly the answer."

That's Twitch's chief product officer, Mike Minton, speaking on a company livestream this week as he explained a new policy: Amazon will now use Twitch creators' videos, clips, and chat logs to train its AI, and everyone is included automatically unless they go into their settings and turn it off.

This arrangement has a name. When something is opt-out, you're in by default and have to act to leave; when it's opt-in, nothing happens unless you choose it. What Minton said out loud is the reason companies reach for the first one: they made it opt-out because almost nobody would willingly choose to opt in.

The same week, Anthropic did a version of the same thing from the other direction, and started adding a watermark to everything Claude writes. A watermark in this case is a hidden pattern woven into the words themselves, invisible to you but detectable later by software that knows to look for it, and it's switched on automatically with no way to turn it off. The company's reason is a new European law requiring AI-generated text to be labeled as such, but whatever the motive, the result is that anything you write with Claude now carries a mark you didn't ask for and can't remove.

One company puts you in and makes you find the exit; the other stamps what you make and never hands you one. Different mechanics, same move: the company sets the default, and it sets it in its own favor, not yours.

You might be thinking, great, now I know that I’m most likely unknowingly being opted into a bunch of stuff, but what can I do about it?

The test is simple: open the tools you actually use and see what you're already signed up for. Then, simply opt out, which in most cases just means turning a toggle from on to off. It should take less than five minutes to do it for all of the AI tools you use. Below are the exact buttons to press for the major ones (I tested them myself when I opted out) ↓

One catch: switching these off usually stops future data, not whatever is already inside a training run. And on Meta's apps in the US, there's no real off switch at all, only an objection form the company doesn't promise to honor.

It’s also important to acknowledge that none of this is really about the settings. You can turn them all off tomorrow during the time it takes you to drink your morning coffee. It’s the fact that you were opted in without being asked, by people who will tell you, on camera, that they did it because asking wouldn't work.

Go check yours and use your autonomy to opt out if you want to. It’s a small thing to do, but it’s your call to make, not theirs.

The AI Collective, Coframe, and Topology Ventures co-hosted a small invite-only dinner in San Francisco for engineers running AI agents in production. Chappy Asel described the table as trading notes on what changes when agents leave the demo and meet real users. A private chef and one of the better views in the city did not hurt.

Thanks to Casey Caruso, founder and managing partner at Topology Ventures, for opening up the space, and to Josh Payne, founder and CEO of Coframe, for pulling the evening together with us. Chappy is collecting names for the next one, so reach out if you want a seat.

Thursday's Lunch & Hack session took on "Decision-Making and Guardrails: Building Reliable AI Agents," walking from planning and memory through evaluations and the guardrails that keep an agent predictable once real users are on the other side of it. Magda Jary led the discussion on how agents actually make decisions and where human judgment belongs in that loop. Carlos Volante, co-founder and CTO of Runtime, ran the hands-on hack.

Adelina Martiniuc, who organizes the series, thanked everyone who showed up, asked questions, and built alongside the group. The full recording is up now, and the next session, "The Future of Work: The New Builder Role," lands on August 27. You can also vote on what gets covered after that at roryplans.ai/Lunch, where Agent Skills and agent orchestration are currently leading the queue.

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.