Weekly AI Highlights: August 8-14, 2026
What Small Business Owners Need to Know Right Now
Hey Reader,
On Saturday I hosted a small dinner party that I swear I thought would kill me.
I used to do this kind of thing all the time but that was before I was vegan.
It’s not that you can’t make delicious vegan food. But the truth is, it takes a lot more prep, and if you are cooking for non-vegans, well, it just adds that much more pressure.
So I started at the farmers market at 8 AM on Saturday buying all the ingredients. At 6:3 pm,, when the guests knocked at the door, I was still cooking! Still wearing my ratty blue jean shorts and sweaty gray T-shirt.
These are the kind of scenarios that happen in my nightmares.
Oh, also in a recurring nightmare, my house is a disaster and the house keeper calls at the last minute to say she can’t make it.
And that happened on Saturday, too.
Good thing the guests are all good friends because as far as I know, they still like me.
Here was the menu (100% AI created, except for the crab cakes).
Speaking of AI…
Claude started putting an invisible mark on everything it writes, and the internet lost its mind about it. Claude quietly moved into your browser sidebar in a way that actually changes what you can hand it. And ChatGPT is now keeping a running record of what you do on your computer all day.
Let's get into it.
1. Claude Is Now Putting an Invisible Mark on the Text It Writes
Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. That is a voluntary agreement about labeling AI content so people can tell where it came from. Here is what it means in practice.
Claude models launched on or after August 2, 2026 now mark what they generate. Older Claude models are being brought into it during a transition period.
There are two kinds of marks. For text, Claude weaves in a watermark you cannot see. It does not change the meaning or the quality of the writing. Because it is baked into the words themselves, it travels with the text when you copy and paste it, and it can survive *some* editing. For files like images, Claude attaches signed provenance metadata. Provenance metadata is just a record of where a file came from and what has happened to it since, using an industry standard called C2PA.
This applies everywhere Claude is offered, worldwide. Not just in Europe. And it covers Claude, Claude Code, Claude Cowork, Claude Tag, and the API.
Now the part that matters most, and Anthropic says this themselves in the article: a detected mark is a signal, not a verdict. Finding a Claude mark does not prove Claude wrote something, because you might have written it yourself and only used Claude to proofread or translate it. And not finding a mark does not prove a human wrote it, because heavy editing, paraphrasing, or a short passage can wipe out the signal.
Source Link: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
MY TAKE
I know the first reaction is going to be "wait, can people tell now?" What I keep coming back to is what Nancy Myrland posted in our AI Club. She pointed out that the more useful question is not "can anyone tell I used AI," it's "do I understand what I'm putting my name on, have I verified what needs to be verified, and am I willing to take responsibility for it?" That is the standard either way, watermark or no watermark. And as far as watermarking in general…I’m all for it. I’m for regulation and transparency. I want for all of us to be able to know what’s real and what isn’t.
2. Claude Cowork Now Runs in Your Chrome Sidebar
Two things to know first. Claude in Chrome is a browser extension that lets Claude see the page you are on and act in it, clicking links, typing, moving between pages, and filling out forms, using the logins you already have. Cowork is the Claude that takes on a whole multi-step job and hands you the finished thing.
Until this week those were two separate worlds. Work you did in the browser sidebar stayed in the browser. Now it is one session. Your conversations save to your history, the skills and connectors you already set up work in the browser with nothing new to configure, and a task you start in a tab can be finished later in the desktop, web, or mobile app.
Here is where it really rocks. Plenty of your tools connect to Claude directly, but the ones that do not are usually the ones eating your time. Vendor portals. Legacy systems. That internal dashboard with the terrible export button. Claude can now work in those through the browser, because you are already logged in.
Anthropic's own example: you are building a budget and need invoice amounts from several vendor portals. Claude opens the tabs, reads each invoice, and builds the spreadsheet. Then you pick the same session up on your desktop to pull in last month's budget and ask what changed.
This is live on Max and Team today and rolling out to Pro over the coming weeks. On Enterprise it is off until an admin turns it on.
Two limits worth knowing. You still need the desktop app to work with files on your own computer. And it does not run on other Chromium browsers or on mobile yet.
On safety, Anthropic is direct about the risk: people hide instructions inside web pages and emails that can redirect Claude to do something you never asked for. In "automatically approve" mode Claude works straight through, but before anything consequential like submitting a form or sending a message, a separate check compares the action against what you originally asked for and blocks it if it does not match. Claude still stops and asks before purchases or sharing personal data. Anthropic says plainly that this reduces the risk without eliminating it, and recommends starting on sites you trust.
Source Link: https://claude.com/blog/cowork-chrome-side-panel
Safety guide: https://support.claude.com/en/articles/12902428-use-claude-in-chrome-safely
MY TAKE
The part I care about is that the session follows me now. Starting something in a browser tab and finishing it on my desktop without re-explaining any of it is the thing I have wanted since the beginning. I would still start somewhere low stakes. Not your bank. (I just wish this was available on Monday, because I literally spent the better part of today trying to find a chat I had with Claude, only to realize I had it in the plugin/sidebar, and it’s gone! 🤦🏻♀️)
3. ChatGPT Can Now Remember What You Did on Your Computer
OpenAI released a feature called Computer History in the ChatGPT desktop app. It turns your activity across your apps and websites into a timeline ChatGPT can look back on, so you stop re-explaining what you were working on.
You can ask it plain questions. “What was I working on before my last break?” “Where is that proposal document I had open earlier?” “Give me a list of what I worked on today and where each thing stands.” (That last one!)
There is a timeline view grouped by day and time. When it notices you doing the same sequence over and over, it offers to turn that into a reusable skill or a scheduled automation.
It records what you clicked, typed, and switched between. It does not take screenshots, record your screen, capture your microphone, or listen to system audio. That is the real change from Chronicle, the earlier version, which did use screenshots. Private browsing is never included, and the temporary activity files are deleted after 48 hours.
You decide what feeds it. You can exclude specific apps and websites, or flip it around and allow only the ones you pick. You can pause and resume it from the menu bar at the top of your screen, and you can delete the last ten minutes, the last hour, the last day, or everything.
Who can use it? Mac only, in the desktop app. Pro, Business, and Enterprise plans. It is off until you turn it on, and it needs Memories turned on to work. On Business and Enterprise an admin has to grant access before you can even opt in. It is not available in the UK, Switzerland, or the European Economic Area.
Two cautions worth reading twice. OpenAI says this raises the risk of prompt injection, meaning a website with hidden instructions on it could trick ChatGPT into doing something you never asked for. And the memory files it saves on your Mac are not encrypted, so another program running under your account could potentially read them.
Source Link: https://learn.chatgpt.com/docs/customization/computer-history
MY TAKE
The timeline is the part I want. Half my week is spent trying to remember where I put something. What I am sitting with is the line in their own documentation telling you to turn it off during conversations with other people unless they have agreed to it. That is a real thing to think about before you switch this on.
PROMPT OF THE WEEK
Turn a Task You Hate Into a Workflow You Can Reuse
Most of us use AI one task at a time. Ask, get an answer, close the tab. Next month, same task, start over.
There is a better move. Before you ask AI to do the work, have it design the workflow. You give it the goal and the context. It maps out the steps. You decide where it has to stop and check with you. Then you save the whole thing and run it again next month without rebuilding it.
Three rules keep this from going sideways. Start with one bottleneck instead of automating the whole job at once. Give it the goal and let it work out the steps, rather than spelling out every move yourself. And stay in the loop anywhere a mistake would cost you real money or a client.
What it is for: taking one recurring task off your plate permanently instead of redoing it every month.
When to use it: on the task you dread. Monthly invoicing, client onboarding, pulling your numbers together, the follow-up sequence you never send.
What you get back: a written workflow with clear steps, a check at each one, and marked spots where the AI has to stop and get your approval.
One thing to notice: the prompt tells the AI not to start yet. That is the whole trick. You are asking it to design the system before it does any work.
Help me design a reusable workflow for this task: [TASK]
Do not do the task yet. Design the workflow first.
1. Define the goal and how we will know it is finished.
2. Lay out the steps in order, from start to done.
3. For each step, tell me: what information you need, what the output should be, and how we will know that
step actually worked.
4. Tell me what happens when a step fails. What gets retried, and what gets handed back to me.
5. Flag every step where you should stop and get my approval before continuing.
6. Tell me which single step is costing me the most time, so we automate that one first.
Keep anything manual that we cannot check reliably yet.
After I approve the workflow, walk me through one real
test run and fix what breaks.
|
Then save it. Once it works, paste the finished workflow into a Claude Project or a Custom GPT so you can run it again without rebuilding it.
Source: The Neuron
|
Comments
If any of this made you think differently about how you want to work with AI, hit reply and let me know. I read every message. Your questions shape what I cover next.
Have a great weekend,
Founder, Front Row AI Club
PS. Was this newsletter forwarded to you? Sign up to get it in your inbox.
You can find my previous newsletter articles here: The Front Row AI Report
© 2026, Front Row AI Club. All rights reserved.