Can my employer tell you used Claude at work?

The watermark is the least of it

Anthropic's marking scheme applies to models launched on or after 2 August 2026, and detection needs the key Anthropic holds. There is no published detector. Your employer cannot run one, cannot buy one, and cannot ask Turnitin for one — Turnitin scores writing style and has no access to any provider's key.

So if the question is "will the watermark give me away at work", the answer today is no. The realistic exposure is somewhere else entirely, and it is much less exotic.

What a managed device actually sees

**Network and proxy logs.** On a corporate network or VPN, requests to claude.ai are visible in the same logs as every other site. This requires no AI-specific tooling.

**Endpoint and DLP software.** Data-loss-prevention tooling is deployed specifically to notice large amounts of company text being pasted into an external web form — which is exactly what using a chatbot with work content looks like. Some of it is configured to block that outright.

**Browser management.** A managed Chrome profile can report extensions, history and sign-ins to an administrator.

None of this is new and none of it is about watermarks. If your workplace can tell, this is almost certainly how.

What the document itself carries

**Version history.** Google Docs, Microsoft 365 and Confluence all keep revision timelines. A document that goes from empty to two thousand finished words in one save looks different from one written over an afternoon, and anyone with edit access can open that history.

**Pasted HTML.** Copying rendered output from a chat interface carries markup — including provider-specific class names — into Word, Outlook, Confluence and Google Docs. This is checkable by anyone who opens the source, and it is the single most concrete artifact in normal office use.

**Invisible characters.** Zero-width characters and non-standard spaces survive most paste routes and show up in search, diffs and exports.

Our checker reports exactly these — what is actually present in a passage, including characters that render as nothing — so you know what a block of text contains before it goes into a company system.

The part worth saying plainly

In a great many workplaces this is not a secret to keep. Plenty of employers now encourage AI use outright, and the risk in those places is not detection but pasting confidential material into an external service — which is a real policy problem regardless of who wrote the words.

Where a policy does restrict it, the durable answer is to check the policy rather than to out-engineer the monitoring. Company devices are administered by people whose job is visibility, and that contest is not winnable.

Questions

Can my employer detect the Claude watermark?

No. Detection requires Anthropic's secret key, and no public detector exists. No employer, vendor or IT department can currently verify a watermark in a passage of text.

Would using Claude on my personal phone avoid all of this?

It avoids the device-side monitoring, because your own phone is not reporting to your employer's admin console. It does not change what the finished document carries — version history and pasted markup are properties of the document, not of the machine you drafted on.

Is pasting work content into a chatbot a problem even if nobody detects it?

Often yes, and this is the more serious issue. Many organisations classify it as sending company data to a third party, which can matter contractually and legally whatever the AI policy says.

Does an AI detector prove I used AI at work?

No. Detectors estimate from style and are unreliable in both directions — they flag human writing and miss machine writing. A score is not evidence of what happened.