The thing gets its own page
OpenAI’s canvas puts the document beside the chat, four months after Anthropic’s Artifacts did the same. Conversation is a good way to ask for a change and a bad place to keep the thing being changed.
On October 3, OpenAI released canvas in beta. When you’re writing or coding, the work opens in its own pane beside the chat. You can select a paragraph and ask for a change to just that part, drag a slider to make the whole thing shorter, or ask for a code review whose suggestions you accept, edit or decline one at a time. Daniel Levine, a product manager at OpenAI, told TechCrunch: “This is just a more natural interface for collaborating with ChatGPT.”
It’s the same layout Anthropic shipped as Artifacts in June. Two of the biggest AI products have now landed on the same answer within four months, which tells me the problem demands it.
Where the work lives
Conversation is a good way to ask for a change and a bad place to keep the thing you’re changing.
Until now, a long piece of writing in ChatGPT lived in the scrollback. Every revision was a new copy further down the page. To compare two versions you scrolled between them, and to change one sentence you asked for the whole thing again and hoped nothing else moved. The work had no address. It was whatever the latest message said it was.
Canvas gives the work a place of its own. The chat becomes the requests, and the pane is the thing being changed.
Make the object the page
I made the same move years ago on Uber’s experimentation platform. The old tool was two long lists of flags and experiments, and everything about a flag was a row. I threw out the lists and made each flag its own page, so the thing you were changing was the thing you were looking at.
The principle carries to AI tools directly. Whatever the person is making should be an object with its own place, its own history, and edits that land on the part you pointed at.
What the pane still owes
Canvas has the place. Next it needs the history. Writers and engineers will want:
- A visible change after every edit, so they see what the model touched before they accept it.
- Versions they can name and go back to.
- Edits that stay inside the selection, every time.
The code review already works this way, with suggestions you approve or decline. The writing side should too. Once people trust that a small request makes a small change, they’ll start using these tools for the work that matters to them.