AI Content Detection for Small Business: Own Your Voice
AI Content Detection for Small Business: Own Your Voice
The AI you write with now signs its own work.
On 11 August 2026, Anthropic said it would embed a machine-readable watermark into text generated by new Claude models, worldwide. It is not a visible label, and it is not metadata you strip out in Word. It is a signature woven into the words themselves, one which travels with the text when you copy and paste it. AI content detection for small business has stopped being a third-party guessing game and become something built into the tools you already use.
Here is what has shipped, what it proves, and the seven things worth doing this month.
The short answer
Two of the three big writing models now mark their own output. Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, and the mark "will travel with the text when it's copied and pasted elsewhere, and may persist through some editing". Google DeepMind's SynthID-Text has run inside Gemini since it was published in Nature in October 2024.
Nobody outside those companies has a working detector yet. Anthropic says it will publish detection details in forthcoming technical documentation. Plan for the day it arrives, not for the day it does not.
What a watermark and a content credential mean
Two different things, often muddled together.
A text watermark is a statistical signal built into the words as the model picks them. SynthID-Text runs a small tournament between candidate next-words and favours the ones scoring higher under a secret function. No extra character, no hidden symbol, nothing to find by eye.
A content credential is signed metadata attached to a file. The standard is C2PA, described by the coalition as "tamper-resident metadata that can be attached to digital content, showing how and when the content was created or modified". Anthropic attaches it to supported SVG, PNG and JPG files. OpenAI has attached Content Credentials to DALL·E images since 2024.
Text carries an invisible signature inside the writing. Files carry a signed record on the outside.
What AI content detection for small business looks like now
| Tool | Text | Images and files | Visible mark |
|---|---|---|---|
| Claude (Anthropic) | Imperceptible watermark in models launched on or after 2 Aug 2026, worldwide, across the app, API, Claude Code, Claude Cowork and Claude Tag | Signed C2PA metadata on supported SVG, PNG and JPG | None |
| Gemini (Google) | SynthID-Text, live in Gemini and Gemini Advanced | SynthID plus C2PA on Nano Banana, Omni and Lyria output | Optional since mid-August 2026 |
| ChatGPT (OpenAI) | Nothing announced | Content Credentials on DALL·E images since 2024 | None |
One change matters for anyone making images. Google made the visible Gemini watermark a toggle in mid-August 2026. Switching off the visible mark has no effect on the invisible SynthID watermark or the C2PA metadata. Both stay on every file, and you are able to check any file by asking Gemini directly, or through About this image, AI Mode and Lens.
None of this came out of nowhere. Transparency obligations under the EU AI Act and California's AI Transparency Act both took effect on 2 August 2026, and they put the marking duty on the model providers.
The platforms are hunting too
On 30 July 2026, LinkedIn shipped a "seems like AI slop" report button. Any reader is now able to flag your post. LinkedIn says the reports give it "a source of signal" for tuning its models, and it swapped its "enhance your post" writing feature for a proofreading tool. Chief Product Officer Hari Srinivasan put it plainly: "People come to LinkedIn to connect with real people and share their real perspectives, ideas, and expertise."
Worth being accurate, because plenty of posts have overstated this. TechCrunch's report does not say a flagged post loses reach. Treat the button as a signal about what the feed will reward, not as a penalty already landing on your account.
The reason LinkedIn built it sits in the data. Pangram Labs scanned 1,002,627 posts across LinkedIn, Medium, Substack, X and Reddit from 24 April 2026, using posts over 50 words. Across the whole dataset, 25.72% of long-form posts came back as fully AI-generated. On LinkedIn it was over 40%, and LinkedIn produced 62% of all the flagged AI content while making up around a third of the sample.
Four in ten long-form posts in your feed were written entirely by a machine. Standing out there is not a high bar any more.
What detection does not prove
Anthropic says a detected mark only shows the content "may have been processed by Claude". It does not prove Claude wrote the ideas. The reverse holds too: "Lack of a detected mark doesn't mean the content wasn't AI-generated."
SynthID-Text weakens under paraphrasing, translation and heavy editing, and it only works across providers who adopt the same scheme. Open-source models running on someone's own hardware sit outside it entirely.
This is not a lie detector. It is a provenance signal with soft edges. Anyone selling you a service built on catching people out is overselling.
Seven things worth doing this month
1. Write down the three things only you know. The pattern you see in clients nobody else names, the mistake you made in year two, the number from your own books. Those go into every piece.
2. Stop giving the AI a blank prompt. Feed it your notes, your voice memo, your last client call. A model with your material writes something closer to you.
3. Read every draft out loud. Cut any line you would not say to a client sitting across from you. This one habit fixes most of it.
4. Check what your images carry. The visible Gemini watermark is optional now, and SynthID plus Content Credentials stay on the file either way.
5. Say how you work, once, in plain language. Nobody is asking a Christchurch coach to label a newsletter. An honest line about where AI helps and where you do the thinking builds more trust than silence.
6. Write your one-page AI policy. What you use it for, what you never use it for, who checks the output before it goes live.
7. Post less on LinkedIn and say more. With over 40% of long-form posts fully machine-written, three specific posts a month will out-perform twelve generic ones.
Frequently asked questions
Will a client know I used AI to write my newsletter?
Not today. Anthropic says it will publish detection details later, and there is no public tool where you paste a paragraph and get a verdict. Write as though the tool already exists and the day it ships will not worry you.
Does the watermark change how the writing reads?
No. Anthropic says the mark does not change the meaning, quality or readability of the text. Google ran a live comparison across roughly 20 million Gemini responses and found the thumbs-up rate differed by 0.01% and the thumbs-down rate by 0.02%, both statistically insignificant.
Do I have to disclose my AI use?
The marking duties introduced on 2 August 2026 sit with the model providers. EU rules do place notification duties on deployers in specific cases, including deepfakes and AI-generated text on matters of public interest. A newsletter about your coaching practice is not one of those. Check your own market's rules and your professional body's guidance.
Does LinkedIn punish AI-written posts?
LinkedIn added the report button on 30 July 2026 and says the reports train its models. TechCrunch's reporting does not confirm a reach penalty. What is certain is the direction, and it is not towards rewarding volume.
Key takeaways
- Claude marks its text output worldwide for models launched on or after 2 August 2026, and the mark survives copy and paste.
- Gemini has carried SynthID-Text since 2024, and turning off Google's visible image watermark leaves SynthID and C2PA metadata in place.
- No public detector exists yet for either, though Anthropic has said one is coming.
- Detection is imperfect in both directions, so treat it as provenance, not proof.
- Over 40% of long-form LinkedIn posts are fully AI-generated. Being specific is now the whole advantage.
Where to start
Pick one piece you published last month and rewrite it with your own material in front of you. Your client notes, your numbers, your opinion. Notice how different it sounds. The difference is the thing worth protecting, and no watermark touches it.
If your voice has never been written down properly, the Strategy package covers brand voice, brand visuals, content pillars and a calendar. If the weekly writing is the part eating your time, the Done & Dusted Content Machine researches your week and drafts the blog, the social posts, the Google Business update and the newsletter for you to check and approve. And the AI Policy template covers step six for $97 USD.
Use the tools. Own the thinking. That has always been the deal.
Need some help with this? Let's chat.
