Leon Furze

Canvas is OpenAI’s latest addition to the ChatGPT interface, offering editing tools for text and code. It rolled out shortly after the release of the o1 model and Advanced Voice Mode for Plus subscribers, coinciding perfectly with OpenAI’s latest fundraising round. Although it’s far from perfect, the feature represents an important step for how AI developers are conceptualising the ways people use their tools.

OpenAI’s Naming Conventions

Let’s get this out of the way first: OpenAI is terrible at naming products. The dubious “Open” in their company name notwithstanding, every model and product they have released has had inconsistent and awkward names:

• Generative Pre-trained Transformer 1, 2 and 3 (GPT-1, GPT-2, GPT-3)
• GPT-3.5
• ChatGPT
• GPT-4
• GPT-4o (with the “o” standing for “Omni”)
• GPT-4
• o1-preview (aka “Strawberry”)
• o1-mini

They’ve also used “GPTs” as the name for their miniature custom chatbot applications, which is nonsense since GPT is just an acronym for “Generative Pre-trained Transformer,” and the GPTs app are not actually GPTs themselves…

The image generation model DALL-E at least has a nod to an actual artist, but given the tensions between the arts community and generative AI, that might not be a good thing.

The newest feature, Canvas, seems to have come from nowhere. The name doesn’t really represent the product since in Canvas you don’t start with a blank slate but with the text that you’re planning to edit. It’s also not a canvas in any artistic sense of the word. I’m positive that the people at both Canvas (the learning management system) and Canva (the graphic design software) are unimpressed with the choice of name.

Calling the underlying model “GPT-4o with Canvas” is salt on the wound for users who already can’t tell which model is which. Given the millions of dollars they spend on public relations and advertising, I can only assume that their atrocious naming conventions are an in-joke, and they’re doing it on purpose.

What is Canvas?

Canvas was released to ChatGPT Plus subscribers on October 3rd 2024, and according to the official OpenAI blog “introduces a new way of working together—not just through conversation, but by creating and refining ideas side by side.”   

One interesting feature of Canvas is that it was designed and built partially using synthetic data. Rather than being a fully fleshed-out feature of ChatGPT, it’s more of a proof of concept. In an earlier post, I wrote about how artificial intelligence will impact the user interface of most digital technologies, and I called miniature applications like Canvas “AI widgets.”

OpenAI releasing Canvas as a widget, designed partially with synthetic training data produced by the o1-preview model, is an indication of how future features will be developed and incorporated into the software. It’s also a glimpse of how we, as users, will interact with these tools in the future, creating our own features and tools on the fly.

Canvas is essentially an editing tool that tries to address some of the failings of the main ChatGPT application. If you’ve ever tried to edit text with ChatGPT, you’ll have hit some frustrations. To edit a text, ChatGPT needs to rewrite the entire thing in every prompt, wasting lots of tokens and taking an unnecessary amount of time, whether you’re editing writing or code. Canvas attempts to solve this problem by reproducing the entire text in the right-hand panel and allowing you to edit and make changes to specific parts of the text without reprinting the entire thing.

It also solves a problem with these tools that I recently heard designer and software engineer Patrick Hebron describe as the “discoverability issue of AI.”

The discoverability problem is this: AI applications are so general-purpose that it’s almost impossible to know exactly what they can and can’t do without extensive experimentation. With a traditional word processing application, for example, all of the features are laid out in front of you, either in the ribbon where you’ll find the Home tab and the review features or in the toolbar across the top of the application.

AI applications like ChatGPT, Claude, or Gemini don’t have a toolbar. They don’t have a handy list of features. Canvas addresses this by providing both an interactive toolbar that hovers on the right-hand side of the screen and more explicit prompts for what the model can be used for.

It’s an approach I noted in the article Hands-On with Google NotebookLM, where I commented on the way that Google had built a small application with a clearly defined purpose to help researchers, educators, and students interact with their notes through a Gemini-powered notebook. I firmly believe that this represents the near future of generative AI. We will have large general-purpose applications like ChatGPT and Claude, and they will, in turn, be used to produce these miniaturised offshoots with much more refined and fit-for-purpose qualities.

How to Use Canvas

Canvas is currently a paid feature, so you’ll need a ChatGPT Plus subscription ($20 per month) or a Teams or Enterprise license. You access Canvas by selecting it from the model dropdown at the top left.

With the model active, the main chat window looks the same as ever until you do something that causes Canvas to load. There are a number of ways this can happen. The easiest is to include the instruction “open in canvas,” “editing canvas,” or similar in your prompt. For example:

This works for both writing and code and will open either version of Canvas (I’ll discuss these separately in a moment).

Canvas will also open itself if the model determines that the prompt requires it. This is where that synthetic data from earlier comes in. Part of the training to develop the Canvas feature involved teaching the model when to open Canvas and when not to. You can read more about that training process on OpenAI’s blog post, but essentially, the developers created synthetic data using GPT-4 to fine-tune the Canvas model to open at the appropriate times.

I found this to be less consistent than simply asking for Canvas. Sometimes it will pop up when you don’t need it, and at other times, it will remain silent when you do.

Using Canvas for Writing

If you’re working with text, such as editing a document, writing outlines, or a similar task, Canvas will open a version more suited to writing purposes. This makes the feature set slightly different and focuses on things like length, complexity, and formatting of text.

The text you are editing will appear in full in the right-hand panel, while a smaller left-hand panel (which can be adjusted with the vertical slider) contains something like the standard ChatGPT interface.

From here, there are a number of ways you can interact with the text. You can use the left-hand panel to chat with the document in the same way as a normal ChatGPT interaction. You can also edit the text directly, just like using a word processor, inserting, deleting, and highlighting to replace passages in the right-hand window. If you highlight a passage, you will get a hover-over action allowing you to chat directly with that part of the text. For example, if you want to edit or revise a single sentence, you can highlight that sentence and type the instructions directly.

Whether you update the text via the left or the right panel, you’ll see a screen refresh animation where the model rewrites the specific parts you’ve identified.

You can also use the writing toolbar on the right-hand side for specific tasks. This is where I think OpenAI’s developers have shown the limits of their imagination when it comes to writing and digital tools because these toolbar options are terrible.

Tasks such as summarising or changing the length or complexity of documents are low-hanging fruit for generative AI writing, and though they might be useful in limited contexts, particularly things like cheap and nasty online advertising copy, they’re not particularly useful tools for serious writers in any form, whether that’s marketing, academic, or creative. They demonstrate a very narrow understanding of what writing is and what people who write are looking for in their tools. Nothing here suggests to me that they’ve interacted with professional authors or writers in any context while developing Canvas.

The addition of the add emojis button ✌️ frankly feels like the same kind of “which focus group did this come from?” idiocy offered by Microsoft when it gave us the “Add Sparkle” button in Copilot in Teams.

Here are a few tools I’d rather see included in the writing version of the Canvas editor:

• Check logic and structure
• Suggest tonal changes
• Improve clarity
• Change form (essay to blog to speech to…)
• Fact check
• Suggest counterarguments
• Improve cohesion and internal consistency
• Check narrative structure
• Anything other than add emojis ✌️

Editing Code with Canvas

The developers of Canvas are obviously on much more familiar territory when it comes to editing code, because the code editing side feels fit-for-purpose. Perhaps this is my bias, because I’m an author, not a programmer, and maybe I’m missing some of the nuance of what it takes to create good code, but it seems to me that Canvas’s features for code editing are a lot more thoughtful.

The left and right panels work in much the same way, but the toolbar is significantly different. The features now are replaced with:

• Add comments
• Add logs
• Fix bugs
• Port code to a new language
• Review code

Each of these seems like a useful feature; much more useful than the writing equivalents like “shorten text,” which frankly just seems like an invitation for sloppy summarisation.

When working with code, Canvas will also frequently write and rewrite its own responses, checking its own output before landing on a final version. In contrast, the writing version tends to give its first attempt at the output, which is often less than satisfactory.

Again, maybe I’m just looking at this through a distorted lens as a writer and not a coder, but I’ve seen lots of positive responses online to the code editing features of Canvas, including from software developers and programmers. Perhaps it’s not surprising, given the people who have developed Canvas are programmers and not authors.

Implications for Education

The bigger picture implication of Canvas is that it stands as an example of where I think these companies are going to take generative AI in the near future. OpenAI has demonstrated that they can produce a feature incorporating synthetic data that acts as a small, fit-for-purpose application. I’m sure we will see similar uses from the big AI developers (OpenAI, Anthropic, Google, Meta) in producing miniaturised applications or “AI widgets” built on top of their larger general-purpose models. Anthropic already has this with its “artifacts” feature which in many ways outperforms ChatGPT Canvas. I imagine we will see AI widgets for math, science, research, creative arts, and a host of other disciplines before the end of 2025.

We’ll also see the integration of other features into Canvas. It currently lacks the multimodal capabilities of GPT-4 and can’t be used with Advanced Voice Mode. By mid-2025, we will likely see a multimodal, internet-connected Canvas with voice features.

There are, of course, huge implications for education in these types of apps. Most students and educators are used to using traditional word processors to write everything from reports to short stories, essays to poetry. What does that look like when a word processor is built from the ground up with generative artificial intelligence in mind? Where does the human author end and the AI author begin in this context? Is it a feature that can be turned off, and if it can be turned off, will anyone bother?

Will Microsoft build similar features into Word in the form of an improved Copilot? Almost certainly. In fact, it’s already happening.

How will teachers of writing encourage students to deal with the fear of the blank page and the “shitty first draft” when our word processing app is essentially an AI widget?

What would this look like in those other disciplines? In software design, how much more attention will be given to broader design concepts, problem-solving, and critical and creative thinking when even core coding tools like Visual Studio Code are “AI-first”?

If you teach English, the humanities, any other writing-heavy subject, or a discipline that involves computer programming, I would encourage you to spend $20 and sign up for a single-month subscription to ChatGPT Plus so that you can try out Canvas right now. By the time Canvas inevitably becomes a free feature (following the pattern of all of OpenAI’s features so far), the landscape will have shifted again, but it will shift in that direction. I think you will be well-served to experiment now and think through the implications for your subject area.

Like everything OpenAI releases, Canvas is basically a beta product: untested, unrefined, buggy as hell, and frequently disappointing. But it’s also a powerful and interesting example of the near future of generative AI.​​​​​​​​​​​​​​​​

Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch: