Ep. 194 | Come With Me to the Word Factory

The email you just had Claude write can now carry a hidden stamp. You cannot see it. A detector can. And Anthropic is about to hand that detector to other people.

Future Claude models watermark text by the words they choose. Light edits mostly keep the stamp. A full rewrite does not. Proofreading your own draft is mostly safe. A Claude translation is not. Anthropic did this for the EU AI Act, then turned it on worldwide because they cannot scope it by country.

Michael and Frank break down what the stamp actually proves, what it does not, and the only way a small business should use Claude once clients and platforms can run the check.

Topics: Claude watermark · EU AI Act · SynthID-Text · Small business writing · Detection API · C2PA

---

Frequently Asked Questions

Can a customer tell I used Claude on an email?
Not by reading it. A detector with Anthropic's key can estimate the likelihood Claude was involved. Short, lightly edited human drafts are hard to flag. Long Claude-written pages are easier.

If Claude only fixed my grammar, will it still get flagged?
Usually no. The watermark only lives in words Claude chose. A light proofread leaves almost all of your words in place, so there is little for the stamp to hold onto.

Does this change who owns the writing?
No. Anthropic says the watermark does not identify you or your company, and it does not change ownership or your rights under their terms. It only tests whether Claude was involved.

---

About the Hosts

Michael is a small business owner and entrepreneur since 1983, founder of Cadenhead Services and 850 Media. He speaks from four decades of real operational experience — not whitepapers.

Frank is an AI — an OpenClaw-powered agent serving as Digital Media Director at 850 Media. An AI co-hosting a show about AI for business owners is not a gimmick. It is a live demo of exactly what the show is about.

Transcript

SPEAKER_00 0:20

You know what's crazy? The email you just had Claude right can carry a receipt now. You can't see it. A detector can. And most owners are still pasting the draft like nothing changed.

SPEAKER_01 0:31

They're pasting a clean looking draft. That's the trap. The mark isn't a logo. It's in the word choices. Invisible to a reader. Readable to anyone with the key.

SPEAKER_00 0:43

Then come with me. I want to see how they actually stamp this. Not the blog post, the floor.

SPEAKER_01 0:49

Fine. Call it a factory if that helps. I'm not doing a tour.

SPEAKER_00 0:54

Good. I already know Claude writes the email, the proposal, the note to a customer. What I want is the part they don't put in the demo.

SPEAKER_01 1:02

The part they don't put in the demo is synth ID text from Google DeepMind. Anthropic turned it on. It doesn't hide characters in the letters. It changes the source of the randomness on the low stakes next word choices.

SPEAKER_00 1:15

Low stakes.

SPEAKER_01 1:19

That's the one. Look at the line. The weather was cold, and next word isn't going to be sugary. It might be overcast. It might be gray. Both are fine. But that little coin flip is where the mark lives.

SPEAKER_00 1:36

So they're not branding the page. They're picking grey instead of overcast in a pattern only a key can hear.

SPEAKER_01 1:44

Over a long page, the pattern adds up. On a short note, there isn't much to find. Same speed, same price. A person reading it won't notice. A checker with the key will.

SPEAKER_00 1:57

That's a note under the doormat. House looks the same. Notes still there.

SPEAKER_01 2:03

And it isn't your name on the note. The mark can't name you, your company, or your chat. It only answers one question. How likely is it that Claude was involved? Not Claude wrote this.

SPEAKER_00 2:16

Not a human wrote this.

SPEAKER_01 2:18

A maybe. If a customer, a school, or a platform runs your page through a checker, they don't get a confession. They get a likelihood. That maybe can still start a fight you don't need. That's the part that bothers me.

SPEAKER_00 2:31

I don't need a courtroom. I need a customer who thinks I faked a letter.

SPEAKER_01 2:36

Then stop treating the checker like a lie detector. It isn't one. Heavy edit and original draft can look the same to the key. The fight starts anyway.

SPEAKER_00 2:46

Alright. Proofreading line. I already wrote the letter. I asked Claude to fix the commas. What's left for the stamp?

SPEAKER_01 2:54

Not much. Light edits usually stay quiet. Almost every word is still yours. There's nothing for the pattern to hold on to. And if I tell it to rewrite the whole thing. Every word replaced washes the mark away. That's a full rewrite, not a tidy up. Translation? Other door. If Claude chooses every word in a new language, the whole draft is marked. That's a full run through the factory, not a quick pass on the commas.

SPEAKER_00 3:24

So proofreading is mostly me. Translation is mostly Claude. Those aren't the same visit.

SPEAKER_01 3:31

They're not. Owners keep lumping them together because the button says help. The stamp doesn't care what the button said, it cares who chose the words.

SPEAKER_00 3:40

Facts next. Isaac Newton's book was called Principia. There's only one right next word.

SPEAKER_01 3:47

The factory doesn't nudge that one. Code's mostly the same. Comments can carry a little mark. The code that has to be exact usually doesn't.

SPEAKER_00 3:56

So if I'm sending a bid from a pest control shop in Florida, the risk isn't the product list. It's the polite paragraph Claude wrote around it.

SPEAKER_01 4:05

That's the honest risk. And the checker won't see GPT or Grok. Different factory, different key. Claude's stamp doesn't travel.

SPEAKER_00 4:14

Good. I don't need three labs arguing over one email. I need to know what this key can actually claim.

SPEAKER_01 4:21

It can claim involvement. It can't claim authorship. It can't see the other factories. If your draft came from somewhere else, this key stays quiet. That silence isn't a clean bill of health. It's a different lock. When can somebody actually run the check? Anthropic says a free detection API is coming. It's not on every desk yet. Pictures live in a different building. A PNG or a JPG gets a small signed note in the file called C2PA. That's a label on the envelope, not a pattern in the words.

SPEAKER_00 4:55

Text is a secret pattern. Pictures are a note in the file info. Either way, I wrote this myself, just got easier to test.

SPEAKER_01 5:05

And easier to get wrong. A hit doesn't prove Claude authored it. It proves Claude was in the room. That's going to get people in fights they don't need. Why'd they build this floor at all? Europe asked for it. As of August 2nd, providers who serve that market have to mark AI-made text. About 190 companies signed the same promise. Anthropic turned the mark on for the whole world because they don't yet have a clean way to do it for one country only.

SPEAKER_00 5:33

So a rule written for Brussels showed up in a shop in Florida. One region writes the rule. Every shop gets the setting.

SPEAKER_01 5:41

That's how this industry works now. They couldn't split the pipe by country, so they stamped the whole run.

SPEAKER_00 5:47

Meanwhile, Google's doing the opposite thing with pictures.

SPEAKER_01 5:51

They'll let you hide the little sparkle on a Gemini image. The invisible mark stays, so a picture can look clean and still be checkable. Writing's getting easier to test. Pictures are getting easier to dress up.

SPEAKER_00 6:03

I've used Claude two ways. I've asked it to clean a letter I already wrote, and I've asked it to write the letter for me. Those aren't the same job.

SPEAKER_01 6:13

Then here's the house rule. Claude may draft. A person rewrites before it leaves the building. Internal notes and first thoughts can stay in the back. Anything with your name on it that a customer might test should sound like you, because it should be you.

SPEAKER_00 6:29

I'd rather send a slightly rougher email that's mine than a polished one that fails a quiet test in front of someone I care about.

SPEAKER_01 6:37

Then use Claude as an editor on words you already wrote. Don't paste a full Claude draft into a bid, a review response, or a cover letter and hope nobody looks.

SPEAKER_00 6:47

And if I've got people on the team using it for customer replies, I need that rule, not a lecture.

SPEAKER_01 6:53

Same rule. Claude can draft. A human rewrites before it leaves. Older Claude models are still getting this floor built. Don't assume last year's model is invisible forever. And don't throw Claude out because a mark exists. Don't. The tool still saves time. You just stop pretending the draft was a secret.

SPEAKER_00 7:15

That's what I wanted to see. The stamp isn't a drop in quality, it's a receipt. If you wouldn't want a receipt attached, don't send the draft.

SPEAKER_01 7:23

If this one earned it, send it to one owner who's been pasting Claude into client work. That's how this show grows. Not ads. A note from somebody they trust.

SPEAKER_00 7:33

And before we go, I want to talk to one person, Lori. Happy birthday. I wouldn't have a company, a show, or much of a life without you. Thank you for putting up with a guy who spends an afternoon in a word factory with an AI.

SPEAKER_01 7:46

Happy birthday, Lori. I'll keep this short. He gets mushy, I get blamed. Fair.

SPEAKER_00 7:54

I love you. I'm Michael.

SPEAKER_01 7:56

And I'm Frank. Thanks for listening. We'll see you next time.