How to Check If Your Content Has an AI Watermark

Most founders and marketing teams assume that if AI content looks clean and reads naturally, there is no trace of how it was made. That assumption is wrong - at least for content produced by Claude.

Anthropic embeds a watermark called SynthID Text directly into Claude's output at the generation level. It is not a tag you can see, a metadata field you can inspect, or a stylistic tell a human editor would catch. It survives light editing. The artifact this article is about is that watermark - and the free public tool Anthropic built to check for it.

This process is for marketing professionals and marketing engineers who need to know whether a specific piece of text was produced by Claude and still carries its mark. It takes about ten minutes per piece. You do not need a paid account, an API key, or any specialist tooling.

One thing to skip: if you are looking for a general "is this AI-written?" score, this is the wrong tool. Claude's Content Checker tells you whether Claude's specific watermark is present. It does not tell you whether a piece was written with GPT-4, Gemini, or any other model. That is a different question and a different workflow.

Step 1: Retrieve the exact text you want to check

The checker works on raw text. Not a URL, not a PDF, not a screenshot. You need the actual string of characters that appeared in Claude's output - ideally before any significant rewriting happened.

Pull the text from wherever it lives: your CMS draft, a Google Doc, a Notion page, a Slack message. Copy the full body of the piece. Do not paste a title or metadata fields alongside it unless they were part of Claude's original output. If you have multiple sections that were generated separately, check them separately.

Source: [CMS name, e.g. WordPress / Webflow / Notion]
Draft title: [TITLE OF THE PIECE]
Date generated (approx): [DATE]
Editing done since generation: [none / light copy edits / heavy rewrite]
Text to check: [PASTE FULL TEXT HERE]

What you get back:

• A clean copy of the text you intend to submit

• A record of how much editing occurred before checking

• A baseline you can compare against if you run the check multiple times

What to do with it:

• Keep this record. If the checker returns a positive result, you will want to know how much editing the text went through before it was flagged.

• If the text was heavily rewritten, note that now - a negative result on heavily edited content is not the same as a negative result on raw output.

• If you are checking on behalf of someone else (a contributor, an agency), document when you received the text and what state it was in.

What this step will NOT tell you: whether the text was ever run through Claude in the first place. That determination comes from the checker itself.

Step 2: Run the text through Claude's Content Checker

Go to claude.com/check-content. This is Anthropic's free public tool. No login required. Paste the text directly into the input field and submit.

The tool checks for two signals: SynthID Text (a statistical watermark embedded in how Claude selects words during generation) and C2PA Content Credentials (a provenance standard attached to certain media outputs where applicable). For most marketing copy - blog posts, emails, social captions - SynthID Text is the relevant signal.

URL: https://claude.com/check-content
Field: [Paste text directly into the main input box]
No additional parameters required.
Check one piece of content at a time.

What you get back:

• A result indicating whether Claude's watermark was detected

• A confidence indicator tied to the SynthID Text signal

• Where applicable, a C2PA credential status for media attachments

What to do with it:

• Screenshot or save the result alongside your record from Step 1.

• If the result is positive, the text carries Claude's watermark in detectable form. Decide how to handle it based on your publication policy.

• If the result is negative, do not treat that as confirmation the content is human-written. Read Step 3 before drawing any conclusions.

• Run the check on the original draft and the published version separately if both exist - editing can shift the result.

What this step will NOT tell you: whether content was produced by any AI model other than Claude. A clean result here says nothing about GPT-4, Gemini, Mistral, or any other system.

Step 3: Interpret the result correctly

A positive result means Claude's watermark was detected with enough statistical confidence to flag. It does not mean the piece is entirely AI-written - Claude may have been used for one paragraph, a headline, or a specific section. The mark does not tell you how much of the content came from Claude, only that enough watermarked text is present to trigger the signal.

A negative result means one of two things: the content was not produced by Claude, or the watermark degraded below the detection threshold. SynthID Text is durable under light editing but degrades under heavy rewriting - paraphrasing, restructuring, and substantial word substitution can reduce the signal to the point where the checker returns clean. A negative result on a piece that went through significant editing is therefore inconclusive, not exculpatory.

Result: [POSITIVE / NEGATIVE]
Editing level (from Step 1): [none / light / heavy]
Conclusion:
- Positive + any editing level = watermark present, Claude used
- Negative + no editing = Claude likely not used for this text
- Negative + heavy editing = inconclusive, watermark may have degraded
- Negative + unknown editing history = treat as inconclusive

What you get back:

• A calibrated interpretation grid you can apply consistently across content reviews

• A documented rationale for each conclusion you draw

What to do with it:

• Use the editing level from Step 1 to qualify every result. A result without that context is hard to act on.

• For inconclusive cases, escalate or apply a secondary review process - this tool cannot resolve ambiguity created by heavy rewriting.

• Do not communicate a negative result as "verified human-written" to stakeholders. The correct language is "Claude's watermark was not detected."

• If you are building a content policy around this, distinguish between "watermark present" and "AI-assisted" - they are not the same thing, and conflating them creates compliance risk.

What this step will NOT tell you: what percentage of the text came from Claude, or whether the undetected portions were AI-generated by a different model.

Step 4: Document and decide

Checking once is not a process. If you are doing this for compliance, editorial policy, or audit purposes, you need a repeatable record that holds up if someone asks questions later.

Create a simple log. One row per piece checked. The fields below are the minimum. You can store this in a spreadsheet, a Notion database, or a CMS custom field - the format does not matter, but the habit does.

Content title: [TITLE]
URL or draft location: [URL or path]
Date checked: [DATE]
Checker tool: claude.com/check-content
Result: [POSITIVE / NEGATIVE / INCONCLUSIVE]
Editing level at time of check: [none / light / heavy]
Action taken: [published as-is / revised / held / escalated]
Notes: [anything unusual about this piece]

What you get back:

• An audit trail you can reference if a piece is questioned

• Pattern data over time - if a contributor's work consistently returns positive, that is a signal worth acting on

• A defensible record if your organization has regulatory or editorial disclosure requirements

What to do with it:

• Review the log monthly if you are checking contributor content at volume. Patterns matter more than individual results.

• If your organization operates under EU AI Act or similar disclosure obligations and needs higher-confidence or bulk detection access, Anthropic runs a separate gated request process for organizations with a qualifying legal basis - this public tool is the starting point, not the ceiling.

• Share the log format with anyone in your team who runs checks, so results are recorded consistently.

• Set a policy before you start logging, not after. Deciding what to do with a positive result after you have one in hand is harder than deciding in advance.

What this step will NOT tell you: whether your policy is legally sufficient for your jurisdiction. That is a question for counsel, not a content checker.

The whole loop on one page

Step 1: Copy the text, record the editing history, note the source.

Step 2: Paste into claude.com/check-content, submit, save the result.

Step 3: Apply the interpretation grid - positive means detected, negative plus heavy editing means inconclusive, negative plus no editing means not detected.

Step 4: Log the result and action taken.

Run this loop before publication for any piece where origin matters - contributor submissions, agency deliverables, any content that will carry a byline or be published under a disclosure policy.

Cadence: for a small team reviewing contributor work, once per submission is right. For a team auditing existing published content, work backwards from most recent to oldest - the watermark degrades over time with edits and updates, so older content will surface more inconclusives.

Where this breaks

Heavy rewriting defeats the signal. SynthID Text is a statistical mark embedded in Claude's word selection patterns. Substantial paraphrasing, restructuring, or synonym substitution dilutes it. If someone wanted to strip the mark deliberately, they could - the checker is not a forensic tool against an adversarial actor.

It only detects Claude. Content produced by GPT-4, Gemini, Mistral, Llama, or any other model will return a clean result. A negative result is not a general AI clearance. If your policy requires detecting AI assistance broadly and not just Claude's specific mark, this tool does not cover that use case.

Short text is less reliable. The SynthID Text signal is statistical - it needs enough tokens to establish a pattern. Very short pieces (a headline, a two-sentence caption, a brief tagline) may return inconclusive results even if they were Claude-generated. The checker performs better on longer-form content.

The result reflects the moment you check, not the moment it was published. If content was Claude-generated but has been live and iteratively updated for six months, what you check today may be substantially different from what was originally posted. Check early or you risk checking something that no longer resembles the original output.