What Changed Underneath

Quiet weeks are the ones worth reading carefully. Nothing this fortnight had the shape of a launch, which is exactly why most of it got underreported. Here's what moved, and what each one is actually asking you to do.

Manus went through with it. Data generated by affected accounts on or after December 29, 2025 was deleted across the twenty third and twenty fourth, with the restoration portal opening on the twenty fifth. This came out of Chinese regulators ordering Meta's roughly two billion dollar acquisition unwound, and the deletion was part of the settlement rather than any security incident.

Why it matters more than it looks. Almost every risk framework a small business uses assumes vendors fail for ordinary reasons. Money runs out, security gets breached, product gets discontinued. This was none of those. The deletion criterion was which corporate parent owned the company while your work was being created. There's no line item for that on anyone's vendor checklist.

What to do. Take your most recent export from whichever platform you'd hurt most without, and try to actually open it. Not make a new one. Open the one you already have. Most people discover in about four minutes that they've been backing up something they can't read, or that the export is missing the relationships between records rather than the records themselves.

The August 2 story got covered heavily: Anthropic now embeds invisible watermarks in Claude's text output and signed C2PA provenance metadata in generated files, worldwide, no opt out, driven by Article 50 of the EU AI Act.

The part that got skipped is that this applies to models launched on or after August 2. Models released before that date are unmarked, and the company has until December 2, 2026 to add support during the transition window.

Why it matters. For the next three months, "unmarked" and "not machine written" are different things, and nobody outside this newsletter seems to know it. Anthropic's own documentation is clear that unmarked content doesn't imply a human wrote it, and that marked content might be a human's work that a model translated, edited, or summarized. That nuance is going to survive approximately zero contact with an anxious client.

What to do. Turn on version history wherever it's off, keep your source material, and write two sentences per project describing what you used and what you did after. Written as routine practice it's worth something. Written after an accusation it's worth nothing.

Security researcher findings published in early August showed tl;dv had an exposed database covering 181,874 meeting records across 84,312 users, with around a thousand marked as actively recording at any moment. Some of what leaked included government and university meetings. The issue was reportedly first flagged in January and was still reachable in late July.

Separately, a California federal complaint against Granola alleges it recorded a participant without notice or consent, and used meeting content by default for commercial purposes unless a user opted out.

Why it matters. These are two different failures pointing at the same thing. One is about who can read your meetings. The other is about who knew the meeting was being read. Both land on the same conclusion: the AI notetaker in your calls is creating a permanent, timestamped, searchable record of conversations that used to leave no artifact, and that record is subject to breach, to subpoena, and to discovery.

What to do. Three things, twenty minutes. Set default sharing on your recordings to private rather than link accessible. Audit who's on your workspace. And put recording language in your calendar invite and say it out loud at the top of the call, because a visible bot helps but is not the same as consent.

The consumer tier converged. ChatGPT Plus, Claude Pro, Gemini and Perplexity all sit near twenty dollars a month, business tiers run twenty five to thirty, Microsoft Copilot sits highest at the top of that range.

Meanwhile the tools underneath quietly repriced. QuickBooks Online moved fifteen to twenty five percent this year with the Plus plan landing around a hundred and ten to a hundred and fifteen a month. Notion put its agent capability on a credit system layered on top of seat pricing. Fathom's free tier now caps AI summaries at five calls a month. Zapier changed how AI steps inside workflows get charged.

Why it matters. None of these individually is a story. Together they're the second wave of AI pricing, where the free and cheap tiers that got everyone hooked start carrying the actual cost of inference. Survey data has the median small business running about five AI tools. Five tools with two repricings a year is a budget line that grew while nobody was watching it.

What to do. Open your card statement and add up every AI subscription. The number is usually higher than the guess by a factor most people find uncomfortable. Then cancel the ones you'd struggle to name a task for. There are always two.

Adoption numbers keep climbing. Survey work this year puts small business AI investment above eighty percent, with the median business running around five tools and planning to add more. Roughly fifty eight percent are using generative tools, about double where it sat in 2023.

And then McKinsey's number, which hasn't moved much: somewhere around six percent of organizations are capturing meaningful value.

Why it matters. That gap is not about tool quality. It's the same gap it was two years ago and it points at the same cause. Businesses buy tools and never rewire the process the tool was supposed to change. The subscription is the easy part. Deciding that a task now runs differently, and that nobody does it the old way anymore, is the part that produces the number.

What to do. Pick one workflow. One. Not a tool, a workflow. Something that happens at least weekly and that you can describe start to finish in four steps. Rebuild that one thing around what you already pay for, wiring it together in Make if the steps cross more than two tools, and measure the hours before and after. That single exercise is worth more than three new subscriptions.

Worth flagging as its own thing because it's the part that turns items two and three from abstract into expensive.

Within about ten days of the watermarking announcement, open source work on reading and testing those signals started circulating. That was predictable and it isn't sinister. What it means practically is that the ability to run a check is about to become free and widely distributed, sitting in browser extensions and hiring tools and marketplace review queues.

Why it matters. The limitations are documented carefully by the people who built the marking. Heavy editing weakens a text watermark. Short passages don't carry enough signal. File metadata gets stripped by format conversion, re saving, or a screenshot. Marked content can be a human's work that a model translated or summarized. Unmarked content can be entirely machine written from an older model.

Every one of those caveats lives in a support article. None of them will live in the browser extension. The gap between what the signal means and what people will treat it as meaning is the actual risk this year, and it's a gap you can only defend against with a record you kept beforehand.

What to do. Nothing technical. Keep drafts, keep sources, and have a plain sentence ready about how you work. The defence isn't a counter tool, it's being the person who can answer the question calmly and produce something.

Step back from the individual stories and there's one shape underneath most of them.

Watermarks in text. Signed C2PA metadata in files, the same standard that camera manufacturers, Adobe, and the BBC already use for images, with more than six thousand members and affiliates in the coalition by earlier this year. OpenAI joined C2PA in May and partnered with Google to embed SynthID in image outputs. Regulators expect layered marking rather than any single technique.

Why it matters. Two years ago the origin of a piece of content was unknowable and everybody argued about it. It's becoming a property of the file. Not perfectly, not immediately, and with real limitations that get ignored. But the direction is one way.

For a small business that mostly means the same thing the last three items meant: the record you keep is becoming more valuable than the claim you make. Drafts, sources, timestamps, and a plain sentence about how you work will do more for you over the next two years than any amount of careful positioning.

If you're reading this on a Saturday and want to close the laptop, here's everything above compressed into things you can actually do.

Open your most recent export. Any platform. The one you already have. Four minutes, and it either reassures you or it doesn't, and both outcomes are worth knowing.

Turn on version history wherever it's currently off. Two minutes. It's free, it's retroactive from today forward, and it's the single cheapest piece of insurance in this entire newsletter.

Set your meeting recordings to private by default. One click.

Add up your AI subscriptions. Five minutes with your card statement. Cancel the two you can't name a task for.

Write two sentences about how you work. What you use, what you do after. Put them somewhere you'll find them.

Fifteen minutes total. That's the whole week's homework, and it's more than most businesses will do about any of this before December.

If you only take one thing: how many hours would it take you to rebuild if your most important platform went dark on Tuesday with no warning?

Not a feeling. A number, written down, including the time to re authenticate everything downstream, which is always the biggest chunk and always the forgotten one.

Most people have never calculated it. The ones who have almost always change something within a week, because the number is worse than the vibe, and a bad number is the only thing that reliably produces action.

Manus users found out this week whether their number was real. Cheaper to find out on a Saturday morning with a coffee.

Jordan

The AI Newsroom is written by Jordan Hale. This issue contains affiliate links to tools I actually use. If you sign up through them I may earn a commission at no extra cost to you.