AI Watermarks: What Happens After We Publish?

In the first post, I looked at whether AI companies are actually watermarking text.

The answer is becoming clearer.

Some are.

Some are not, at least not in ordinary text, based on the official documentation we could find.

And some still haven’t given us a clear public answer.

But that raised a bigger question for me.

What happens after that content leaves the AI platform?

If text has been helped, corrected, rewritten or generated by AI, and that text carries a watermark or another provenance signal, what do third-party platforms actually do with it?

Do they label it?

Do they ignore it?

Do they reduce its reach?

Do search engines rank it differently?

Or is the watermark simply there to show where the content came from?

That is what I wanted to find out next.

First: what the AI companies themselves confirm

Anthropic / Claude

Anthropic officially confirms that future Claude models generate text containing a statistical watermark.

The watermark does not add hidden characters. Instead, Claude slightly changes some of the word choices it makes while generating text, creating a statistical pattern that can later help determine whether Claude was involved.

Anthropic says the watermark:

• does not practically affect the quality of the output,
• does not identify the individual user or organisation,
• does not add extra tokens,
• and is being introduced partly to comply with the EU AI Act.

We’ll come back to the EU AI Act in one of the next posts, because that part of the story deserves a proper look on its own.

Source below ↓

Another useful detail is that Anthropic distinguishes between fully generated text and lighter AI assistance.

When Claude is only proofreading or lightly processing human-written text, the watermark may be much weaker because many of the original human words remain unchanged.

That distinction matters to me because this is much closer to how many people actually use AI.

The idea may be human.

The experience may be human.

The opinion may be human.

AI may simply help turn it into clearer language, restructure it, translate it, or correct it.

Google / Gemini

Google DeepMind confirms that SynthID is used to watermark AI-generated images, audio, video and text.

For text, Google says SynthID changes the probability of which words or tokens are selected during generation, creating a watermark that is not visible to the reader and does not affect the quality of the output.

Google specifically confirms that this applies to text generated through the Gemini app and web experience.

Source below ↓

So in Gemini's case, the answer is clear:

Yes, text watermarking exists.

OpenAI / ChatGPT

OpenAI officially documents provenance signals for images and audio.

Supported images can contain both C2PA Content Credentials and SynthID, while supported OpenAI-generated audio can contain a SynthID watermark. OpenAI also operates a verification tool that currently checks images and audio.

Source below ↓

Source below ↓

The important part is what the documentation does not currently list.

Ordinary ChatGPT text is not included in OpenAI's current table of supported provenance signals.

That does not prove that OpenAI could never introduce text watermarking.

It simply means that, based on the official documentation available now, we cannot say that ordinary ChatGPT text currently carries the same supported provenance signals as images and audio.

xAI / Grok

xAI officially confirms that Grok-generated images and videos include a watermark.

xAI also says there is no setting to remove it and that removing or obscuring provenance signals is prohibited under its Acceptable Use Policy.

Source below ↓

What we still could not find is an official xAI statement confirming that ordinary Grok text carries a statistical or hidden text watermark.

So for Grok text, the answer remains:

Not publicly confirmed.

We have asked for clarification and are still looking.

Then comes the real question: what do third-party platforms do with it?

When we started asking the AI companies about watermarking, one thing became obvious.

Even if an AI provider tells us what its own watermark does, it cannot necessarily tell us what another company does with that signal once the content is published elsewhere.

That means we also need to look at the platforms themselves.

YouTube

YouTube officially confirms that it can read C2PA Content Credentials and use them to carry forward disclosures about how a video was made.

Its help documentation says content can receive an “Altered or synthetic content” disclosure through creator disclosure, YouTube's own AI tools, or valid C2PA credentials.

Source below ↓

What I could not find in YouTube's official documentation is a rule saying that the presence of an AI provenance signal automatically reduces recommendations or distribution.

The documentation describes the system primarily as a transparency mechanism.

So at the moment:

TikTok

TikTok says it is strengthening AI transparency systems and is part of the C2PA ecosystem.

TikTok has also said it is testing improved detection systems aimed at accounts that publish AI-generated spam which crowds out original creators.

Source below ↓

That distinction is important.

TikTok is talking about spam behaviour and low-value AI content, not simply the fact that AI was involved.

We still need to be cautious here.

I have not found an official TikTok statement saying that properly disclosed, useful AI-assisted content is automatically downranked merely because a watermark is present.

So:

LinkedIn

LinkedIn now gives us one of the clearest answers for human-assisted writing.

Its official guidance says AI can be used for things such as:

• refining language,
• proofreading,
• making writing more concise.

But LinkedIn says the content should still reflect the member's own voice, perspective and experience.

It also distinguishes AI-assisted content from what it calls “AI slop”: generic, repetitive, low-value content that lacks substance or a genuine point of view.

Source below ↓

LinkedIn also says AI-assisted content is welcome when it reflects real expertise or experience, while generic content is less likely to be widely distributed.

For me, that is an important distinction.

LinkedIn is not saying:

AI = bad.

It is saying:

Low-value, generic content = less useful.

I still have not found an official LinkedIn statement saying that a hidden text watermark itself triggers downranking.

X

X's official authenticity rules focus on deceptive or harmful manipulated media.

The policy covers media that has been significantly altered or fabricated in a way that changes its meaning and could cause confusion, safety risks or serious harm.

Source below ↓

What I have not found is an official X policy saying that ordinary AI-assisted text is reduced in reach because a watermark exists.

So for X:

Reddit

Reddit's official spam policy explicitly mentions generative AI tools in the context of spam proliferation.

It prohibits repeated or unsolicited mass engagement, mass posting and the use of tools, including generative AI, to facilitate spam.

Source below ↓

Again, the problem being described is spam behaviour, not simply AI assistance.

I found no official Reddit policy stating that a text watermark by itself causes a post to be automatically downranked.

Bluesky

Bluesky's official Community Guidelines describe moderation around safety, deception, abuse, synthetic content and other harmful behaviour.

Bluesky also allows independent labeling systems through its composable moderation architecture.

Source below ↓

I have not found official Bluesky documentation saying it detects AI text watermarks or reduces the visibility of ordinary AI-assisted writing because of them.

So that remains unconfirmed.

Substack

Substack's official Content Guidelines focus on legality, abuse, harmful content, copyright and other policy violations.

Substack reserves the right to hide or remove content that violates those rules.

Source below ↓

I could not find an official Substack policy saying it detects AI text watermarks or reduces newsletter distribution simply because AI assisted with the writing.

Again:

No confirmed watermark-based penalty found.

Quora

This one is still not clear enough.

I do not currently have a reliable primary Quora source confirming how it treats AI text watermarks or whether AI provenance affects ranking or distribution.

So I would rather say:

We don't know yet.

And keep looking.

What about search engines?

This is probably the part that matters most for anyone publishing on their own website.

Google Search

Google's position is actually quite clear.

Google says appropriate use of AI or automation is not against its guidelines.

Its focus is on whether content is:

• useful,
• original,
• reliable,
• created primarily for people,
• and not produced at scale mainly to manipulate rankings.

Source below ↓

Google explicitly says that its focus is on the quality of content rather than how it was produced.

Source below ↓

Google's spam policy does allow it to reduce rankings or remove sites that engage in things like scaled content abuse.

That means generating large amounts of low-value or unoriginal content mainly to manipulate search rankings, regardless of whether humans or AI created it.

Source below ↓

This is probably the clearest answer we have so far for our original concern.

AI assistance itself is not the problem.

Manipulative, low-value content can be.

I have found no Google Search policy stating that the existence of a text watermark itself causes a ranking penalty.

Bing

Microsoft's official Bing Webmaster Guidelines are also useful.

Bing says large-scale automatically generated content without oversight, quality control or editorial review may be excluded from indexing.

It also warns against scraped content, artificial language, manipulation and low-value material.

Source below ↓

But again, I did not find an official Bing rule saying:

“AI watermark detected = ranking penalty.”

The focus is on quality, originality, oversight and manipulation.

DuckDuckGo

DuckDuckGo is less clear.

I have not found official DuckDuckGo documentation explaining whether it detects third-party text watermarks such as SynthID or Claude's statistical watermark, or whether such signals affect ranking.

So I do not want to pretend we know.

For now:

DuckDuckGo: unknown.

We are still looking.

And the same applies to several other search engines where there is no clear public documentation.

So where does this leave us?

What we know today is more nuanced than the original fear.

AI companies are increasingly introducing ways to identify AI-generated content.

And I actually think that can be a good thing.

If a photograph is generated by AI, it makes sense to be able to know that.

If a video shows something that never happened, knowing that matters.

If a voice was synthetically created, provenance can protect people from being misled.

Even text provenance itself is not necessarily a bad idea.

My question was never really:

“How do we hide that AI was involved?”

My question was:

“If AI helps me correct, restructure or express my own idea, does that technical watermark then work against the content?”

So far, the answer seems to be:

But we have not found broad evidence that useful human-led AI-assisted content is automatically penalized simply because a watermark exists.

Where the official documentation gives us an answer, we will publish it.

Where it doesn't, we will say:

We don't know yet.

And where necessary, we will contact the companies directly and ask.

There is still another part of this story.

Why are companies implementing these watermarking systems now?

One of the biggest pieces behind that question is the EU AI Act.

That deserves its own article.

Because once we understand what the law actually requires, and what it does not require, the rest of this starts to make much more sense.

And I Still Have One Question

After everything we found, I keep coming back to something from the first post.

If a watermark by itself is not proven to damage ranking, reach or visibility, why are we being told that we need to remove it from our text?

And why do watermark-removal tools exist in the first place?

There is another part of that question too.

What do those tools actually do to the text we give them?

If an AI company can create a statistical pattern inside generated text, what happens when a third-party service claims to remove it?

Does it simply rewrite the text?

Does it change something else?

Could it accidentally affect the meaning, quality or structure of the content?

Could it introduce another detectable pattern of its own?

At this point, I don’t know.

And that is exactly why I would be careful.

Based on everything we have found so far, I personally don’t see a clear reason to pay for a service simply to remove an AI text watermark.

Not because I can prove those services are bad.

I can’t.

But because we still haven’t found strong evidence that the watermark itself is causing the problem people are being told to solve.

The official information we have found points much more towards quality, usefulness, spam, deception and policy compliance than towards punishing content simply because AI helped create or edit it.

So for now, my approach is simple:

If the idea is genuine, the information is useful, the content is legal, and AI is helping me communicate it better, I’m not going to panic about the watermark.

And until we understand exactly what watermark-removal tools do, I’d be careful about handing important content to one just because someone says the watermark needs to disappear.

There is still more to investigate.

The EU AI Act is part of that story.

So is the reason companies are introducing these systems globally.

And we still want clearer answers from several platforms about what happens after watermarked content is published.

To be continued.

No noise. Just open talk. No shortcuts.