Artifacts

Anyone working with a chat program based on artificial intelligence normally receives answers as continuous text within the conversation history. That’s practical for short queries, but it becomes confusing as soon as the result gets longer. A ten-page essay or a piece of program code disappears upward out of view after a few more messages. Artifacts solve this problem: such larger results end up in their own window next to the conversation. There they remain in place while the conversation continues, and they can be specifically edited on request. The term comes from English and means roughly “a produced piece of work.”

From conversation log to working document

The difference sounds cosmetic, but it changes the nature of the collaboration. In a pure chat, every follow-up question produces a completely new version of the answer. The history fills up with nearly identical blocks of text, and one has to figure out for oneself which one is the current version. With an Artifact, on the other hand, there is a single document that changes. The conversation next to it becomes the instruction trail, the window becomes the actual result.

One can picture this like a workshop. The chat window is the conversation with the craftsman, the Artifact is the piece on the workbench. You say “make the handle shorter,” and the handle gets shorter — you don’t get a whole new chair delivered every time. It is precisely this feeling of working together on one thing that explains the popularity of the feature.

For companies this is also economically appealing. If a result can be used directly, the detour of copying, pasting, and opening a second program is avoided. Providers such as Anthropic, OpenAI, and Google have therefore all introduced comparable features, sometimes under different names such as Canvas.

What actually happens in the window

Technically, the program decides based on a few rules of thumb whether an answer belongs in the side window. Typical triggers are length, a clearly delimited piece of content, and the intention to reuse the result later. A calculation stays in the chat, a 300-line program file moves into the Artifact. The model’s response contains invisible markers for this purpose that tell the interface: this section is a self-contained document.

When changes are made, usually not everything is rewritten. The model often only specifies which part is to be replaced by which new content. This is faster and cheaper, because less text needs to be generated. Many interfaces additionally save every version, so that one can jump back and forth between versions.

A special feature concerns code for web pages. Such Artifacts are not just displayed but are actually executed directly in a sandboxed environment in the browser. So instead of seeing the code of a calculator, you see a working calculator. The sandboxing is important, because the code comes from a machine and has not been reviewed by anyone.

Typical use cases and their limits

In everyday use, the feature appears wherever something finished is meant to come out: a job application, a table of vocabulary, a diagram, a small learning app. In school and university settings, Artifacts work well for outlines that get rearranged multiple times. In companies, they give rise to prototypes shown to colleagues before developers really get started.

In the news, the term usually comes up when a provider expands its interface. Anthropic introduced Artifacts in 2024 and later allowed them to be shared with others. Observers see this as a trend: chat programs are evolving from information services into working environments. Anyone reporting on AI products today almost always means this kind of side window when using the word.

A common misconception is that Artifacts are stored files in the classic sense. They initially live within the respective conversation and are bound to it. Anyone wanting to keep them must download them or copy them elsewhere. And a second point remains: the fact that a result is neatly formatted in its own window says nothing about its correctness. It still has to be checked.