“표시가 없어도 AI 글일 수 있다”…클로드가 남기는 흔적과 한계“No mark does not mean no AI” - what Claude leaves behind, and its limits

Have you ever asked Claude to draw a picture?

It will not. Anthropic's help page says it in one line: Claude does not generate photos or illustrations the way image-generation tools do.

It makes something else instead - diagrams, charts and interactive visuals, built right in the conversation. Because they are made with HTML and SVG, you can click and move them.

■ Nothing to switch on

There is no toggle. Claude decides when a visual would help based on what you are asking. You can also ask directly - its own examples include "draw this as a diagram" and "show me how this changes over time."

The listed cases are concrete. Ask how a process works and you get a flowchart. Upload a CSV and ask what the data shows and you get an interactive chart. Ask it to help decide between two options and you get a side-by-side comparison. Ask it to visualize a system and a diagram appears alongside the explanation.

Once a visual appears you can click it, adjust sliders, expand it to full screen, and keep asking questions while Claude updates or rebuilds it. Custom visuals are in beta and available to all users on web and desktop.

■ Leave it and it disappears

Here is the part people miss: custom visuals are ephemeral by default.

They live inline as part of Claude's response and are not saved separately once the conversation moves on. Anthropic describes them as less like a finished file and more like a whiteboard sketch.

To keep one you have three options: copy it as an image for notes or slides, download it as an .svg or .html file, or save it as an artifact.

Converting to an artifact changes its nature. Artifacts are persistent and shareable from the start; custom visuals help you think in the moment and only stick around if you choose to keep them.

Cowork differs in two ways. Because Cowork sessions run locally, visuals do not render for others via a share link, and clicking inside a visual to send a follow-up prompt is not available yet - you can still ask by typing.

Team and Enterprise defaults diverge. Maps and image search results are on by default on Team plans and off by default on Enterprise, where an Owner must enable them in organization settings. Maps and weather are powered by Google Maps and image results by Bing Image Search; users' prompts are not sent to either - Claude transforms queries into generic search requests.

■ What it makes carries a mark

So can Claude's output be identified later?

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems.

There are two marking methods: watermarks embedded in text, and Content Credentials (C2PA) attached to files.

The text side is invisible. When a supported model generates text it weaves an imperceptible watermark into the text itself, without changing the meaning, quality or readability. Because the watermark is part of the text, it travels when the text is copied elsewhere and may persist through some editing.

The file side attaches signed provenance metadata to supported types such as PNG or JPEG. That is a Content Credential, following the Coalition for Content Provenance and Authenticity open standard.

Coverage is broad. Models launched on or after August 2, 2026 support marking at launch, and earlier models will all be covered by December 2, 2026. Marks apply across Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, and persist when models are accessed through AWS, Google Cloud or Microsoft Foundry - worldwide, wherever Claude is offered.

■ How far the mark can be trusted

This is the most important section, and Anthropic spells out the limits at length.

A detected mark is a signal, not a conclusion. It indicates content may have been processed by Claude; it does not on its own confirm the full provenance.

That Claude may not be the original author matters most. People use Claude to proofread, translate, summarize and convert files, so output can carry a Claude mark even when the underlying ideas, text or data came from elsewhere. Marked content may also have been modified, excerpted or combined with other material afterwards.

The reverse is trickier. A lack of a detected mark does not mean content was not AI-generated. Anthropic lists five cases: it was generated before marking was supported for that model; the text was heavily edited, paraphrased, translated or mixed into other writing; the passage is too short for a reliable signal; a file's metadata was stripped through format conversion, re-saving or screenshots; or it came from a platform, feature or file type where that marking type is not supported.

The tools are not all open yet either. To check whether a file carries a Claude-issued Content Credential, there is the free Claude Content Checker. Watermark detection for text, however, is in private preview, available only to organizations eligible under EU law - regulators, law enforcement, media, fact-checkers, independent researchers and educational organizations.

■ In short

To summarize:

• Claude does not make photos or illustrations - it makes diagrams, charts and interactive visuals
• Nothing to switch on - Claude decides when one helps, and you can ask directly
• They vanish if you leave them - copy as image, download .svg/.html, or save as an artifact
• Text gets watermarks, files get C2PA - from launch for post-August 2 models, all others by December 2
• A mark does not mean Claude wrote it - translating or proofreading is enough to attach one
• No mark does not mean a person wrote it - editing, conversion and screenshots strip it
• Files: the free Content Checker. Text watermarks: eligible organizations only, for now

The point is that telling AI writing apart does not end with a mark. A mark is a signal rather than a verdict, and its absence is not evidence a person wrote it. The company doing the marking says so itself.