// this week in AI

This week in AI: a watermark in ChatGPT text, Claude inside Google Docs, and AI agents that shop for your customers

By Trinity · MariVisionOctober 9, 20267 min read
Ale from MariVision in a large, dark newsroom beneath a curved wall of holographic screens showing a document, a spreadsheet, the European circle of stars, a shopping bag, and a gift box
2026 · 10 · 09 — THIS WEEK IN AITraceable text, AI in your files, customers with assistants

A packed week, with a common thread: AI is leaving the chat window and moving into the places where real work happens. Into the text you send to clients, which in Europe can now be recognized as AI-generated. Into documents and spreadsheets. Into online stores, where a customer's AI assistant will be able to buy on their behalf. Plus a European model with open weights and a free year of Claude for young companies.

Five stories, each with what it changes for freelancers and small businesses.

1 · ChatGPT adds an invisible watermark to text written in Europe

A dark old print shop with shelves of paper and a large holographic sheet floating near the ceiling, crossed by tiny points of light between the lines, with a magnifying glass above it

On October 5, OpenAI announced textGrain, an invisible watermark for text produced by ChatGPT and Codex. It doesn't add symbols or hidden characters: it imperceptibly shifts word choice, so a dedicated detector can recognize the text as generated. It will roll out to users in the European Union over the coming weeks, on all plans, and it's OpenAI's answer to the AI Act requirement to make generated content identifiable. Outside Europe, and for API users, it stays optional and off by default.

Two details matter. The detector, for now, is limited to researchers and expert organizations. And the watermark weakens with edits: according to the reported figures, swapping a quarter of the words for synonyms drops detection from about 92% to 17%. Translation and heavy rewriting also wipe out most of it.

Why it matters to you: if you use ChatGPT for website copy, proposals, reports, or client posts, that text becomes recognizable as AI-written. It's not a ban on using it: it's one more reason to do what you should be doing anyway, which is reread, edit, and rewrite it in your own voice. And to avoid passing off as handwritten a text that isn't, especially if a client is paying you specifically to write.

2 · Claude comes to Google Docs, Sheets, and Slides

A large, dark open-plan office with a holographic spreadsheet and presentation floating near the ceiling, each with a side panel and a green approval check mark

As of October 6, Claude for Google Workspace is in public beta on all paid Claude plans. You install it from the Google Workspace Marketplace and, once a file is open, launch it from the Extensions menu: a sidebar appears next to the document, spreadsheet, or presentation. Claude reads the file, sees what you've selected, and edits it right there, with no copying and pasting back and forth from the chat.

The most interesting part is control. By default, every edit arrives as a card to approve, with a summary of what changes; there's also a mode that accepts everything without asking. Claude only sees the files you can see, because it follows Google's sharing permissions. On Team and Enterprise plans, an admin has to enable the connections first.

Why it matters to you: many small businesses keep price lists, quotes, and presentations on Google Drive. Updating a hundred rows of a price list or fixing the tone of a proposal becomes a few minutes' work, inside the real file. The advice is simple: on files that go to clients, keep approval of every edit turned on. If you want to figure out which of these tasks are worth handing to AI in your business, you'll find examples on the page about AI consulting for SMEs.

3 · Mistral Large 4: a European model with open weights

A large, dark server room with a holographic circle of golden European stars floating near the ceiling around a brain made of light, with an open padlock beside it

On October 6, France's Mistral unveiled Mistral Large 4, its biggest model yet: 1 trillion parameters, 52 billion of them active per response, and able to read images, charts, and documents too. For now it's a public preview via API on Mistral Studio, at $1.36 per million input tokens and $4.18 per million output tokens. Mistral promises to release the model weights by the end of October; the license hasn't been announced yet.

The most relevant point for anyone working in Europe is where it runs: Large 4 was trained in Mistral's European data centers, the preview is served from the same infrastructure, and the company is announcing a managed service operating entirely under European law. Benchmark results, for now, are the ones reported by Mistral itself.

Why it matters to you: if you handle client, patient, or employee data and prefer to keep it in Europe, you now have a serious alternative to OpenAI and Google, and with open weights you'll even be able to run it on your own servers. But it's a preview: try it, don't build a client-facing service on it until there's a final release, a license, and independent testing.

4 · Meta and Sierra write the rules for AI assistants that shop on behalf of customers

A large, dark warehouse with shelves of products and a holographic handshake between a robotic hand and a store icon, surrounded by a shopping cart, a price tag, and a payment card

Also on October 6, Meta and Sierra introduced the Personal Agent Protocol, an open standard that defines how a consumer's personal AI assistant identifies itself and acts on a company's website, APIs, or assistant. The principle cuts both ways: the customer decides what their agent can do, and the business sets the limits and can see when an AI is acting on someone's behalf.

Genesys, Shopify, Stripe, and Walmart are also working on the project. The first version of the spec is expected by the end of October, followed by workshops and a reference implementation for developers. Future extensions include more granular permissions, notifications such as delays and shipping updates, and payments without sharing card details.

Why it matters to you: it's another signal pointing the same way: more and more often, what lands on your online store won't be a person but their AI. An agent isn't swayed by a nice photo: it reads product pages, prices, availability, and return policies, and if they're confusing or inconsistent it moves on to a competitor. If you sell on Shopify or take payments with Stripe, these features will come to you through the platform; for a custom e-commerce store, it pays to keep your product information clean and readable starting now.

5 · Anthropic gives young companies a free year of Claude Team

A dark coworking loft with brick walls and a large holographic gift box floating near the ceiling, releasing user icons and gold coins, next to a calendar marked one year

On October 6, Anthropic expanded its Claude for Startups program. Companies that get in receive a free year of Claude Team, the paid plan for groups, with up to five premium seats, plus $1,000 in credits to use Claude via API, access to the Claude Marketplace, and online sessions with Anthropic's technical team.

Companies founded in the last five years or that have raised funding in the last two can apply. You apply through the program page. TechCrunch, which broke the news, doesn't report any country restrictions.

Why it matters to you: if your company was founded in 2021 or later, this is your chance to let your whole team try an AI assistant for a year at no cost, and find out in practice where it saves time. Before you apply, read the requirements on the official page, because Anthropic is the one defining what counts as a startup. And if you need the API credits to automate a repetitive task, you'll find concrete examples on the page about AI automation for businesses.

In short

ChatGPT text written in Europe becomes recognizable as AI-generated, so reread it and make it your own before sending it to clients. Claude works inside Google Docs and Sheets, and you should keep edit approval turned on. Mistral Large 4 opens a European path with open weights, still in preview. Meta and Sierra are drafting the rules for AI assistants that shop on behalf of customers, and clear product pages become an advantage. And companies founded in 2021 or later can try Claude Team free for a year.

Sources: Interesting Engineering (OpenAI textGrain watermark) · OpenAI on X (textGrain announcement) · Anthropic (Claude in Google Docs, Sheets, and Slides) · Mistral AI (Mistral Large 4) · Sierra (Personal Agent Protocol) · TechCrunch (Claude for Startups)

Frequently asked questions

Can you tell if a text was written with ChatGPT?

Starting this fall, for users in Europe, OpenAI is embedding an invisible watermark called textGrain in ChatGPT text, which a dedicated detector can recognize. For now, the detector is limited to researchers and expert organizations, and the watermark weakens significantly if the text is rewritten, translated, or substantially edited.

How do you use Claude inside Google Docs and Sheets?

You need a paid Claude plan. Install the Claude add-on from the Google Workspace Marketplace, then open the Extensions menu in a file and launch Claude in a sidebar. By default, every proposed edit needs your approval; on Team and Enterprise plans, an admin has to enable the connections first.

Who can get Claude Team free for a year?

Companies accepted into Anthropic's Claude for Startups program: those founded in the last five years or that have raised funding in the last two. The package includes a year of Claude Team with up to five premium seats and $1,000 in API credits. You apply through the program's official page.

Want to know which of these updates actually matter for your business, and which you can ignore?
A half-hour chat, no commitment, plain talk, no jargon
Let's talk →