Shortly after OpenAI announced a significant $6.6 billion funding round, the company has unveiled a substantial upgrade to the interface for ChatGPT.
In what may signal OpenAI’s recognition of the potential for its groundbreaking chatbot to move beyond mere question-and-answer interactions, the new beta feature introduces an editable canvas. This feature will appear in a window next to the standard ChatGPT chat box.
“Our primary goal is to create a more effective way for users to collaborate with ChatGPT in writing and coding tasks,” explained Daniel Levine, a product lead at OpenAI overseeing the canvas feature. Starting today, Canvas will be available in beta to subscribers of ChatGPT Plus and Team plans, with Enterprise and Edu subscribers expected to gain access next week. Currently, the feature is fully operational on desktop devices, while mobile users can only view canvas projects.
During a pre-launch demonstration for WIRED, Levine emphasized the aspiration for a more “natural” interaction between humans and AI through this new capability. His team utilized synthetic data produced by OpenAI’s latest model, o1-preview, to enhance the functionality of GPT-4o’s model, allowing it to effectively recognize when to activate features and how to modify documents as needed.
With the canvas feature, ChatGPT can initiate a draft for a project that you can refine together, or you can provide an existing draft for critique. Levine began his experience by asking ChatGPT to utilize canvas for drafting an important email. He then highlighted certain paragraphs and requested targeted modifications. The generative AI tool can insert comments in the canvas to suggest improvements or can even alter the document’s content directly.
Launching a canvas is straightforward: Simply include a phrase such as “use canvas …” or “start a canvas …” in your prompt, and a new window will pop up. In certain writing requests, like prompting for a draft blog post, the GPT-4o model is programmed to automatically start a canvas if it determines that it would be useful. For programming tasks, the model initiates a canvas only when explicitly asked.
While both writing and coding modes allow for in-line editing requests, the user interface for canvas is split into two sets of shortcuts—one tailored for those engaged in AI-assisted writing and the other for coders. During the demonstration, Levine illustrated how the writer’s shortcuts could be employed to shorten text or apply a “final polish” to a draft. He also used one of the more playful shortcuts to insert a variety of random emojis. On the programming side, ChatGPT can incorporate logs, comments, and work on troubleshooting issues within a canvas.
ChatGPT maintains multiple versions of the canvas as you make revisions, allowing you to revert to previous versions if you favor them. Writers concerned about their uploads potentially training OpenAI’s model should navigate to their user settings and ensure that “model training” is disabled.
By allowing ChatGPT to make both edits and suggestions, OpenAI is creating a blurred distinction between authorship and the curation of words. As someone who collaborates with professional editors regularly, I have doubts that the canvas beta will deliver the same sharp insights and thorough guidance that human editors provide. However, for individuals lacking easy access to human writing collaborators, I can understand how receiving artificial notes on composition regarding structure and content could be advantageous.
It is important to mention that three individuals who were identified as “supporting leadership” for the canvas project are no longer part of the organization. John Schulman, a former co-lead and co-founder, left in August to join Anthropic, a competing AI firm. Moreover, former chief technology officer Mira Murati and research vice president Barret Zoph both resigned from their roles just a week prior to this launch. During a press event at the OpenAI office after these departures, the current chief product officer Kevin Weil reiterated the company’s dedication to continue rolling out software updates.
“I believe that 2025 will be the year when agentic systems truly enter the mainstream,” he stated. The concept of an AI “agent” that can not only assist you with software tasks but also be sufficiently agile to traverse the digital landscape to perform actions on your behalf represents both the recent history and an anticipated future of generative AI.
Last year, WIRED reported on ChatGPT’s plug-ins which users could utilize for various tasks, such as booking flights through Expedia or making dining reservations via OpenTable—this was seen as a significant move toward more “agentic” AI solutions. Nevertheless, the plug-ins were subsequently discontinued, leading to the introduction of more limited custom GPT chatbots as their replacement.
With that context, the beta version of the canvas seems to be yet another effort to enhance AI models by granting them additional decision-making capabilities, resulting in unexpected outcomes. In a demonstration by WIRED, Levine pointed out a specific area of the canvas and requested an alteration. In response, ChatGPT made an inline adjustment at the bottom of the canvas, away from where he had highlighted. “What’s fascinating is that when you select a section, it often makes a change in that area,” he explains. “However, ChatGPT has the flexibility to choose where to make edits.”
The nearest alternative to OpenAI’s canvas tool currently available is likely Google’s Gemini integration, which allows users to leverage generative AI within Docs or explore Anthropic’s Artifacts tool. While chatbots are not disappearing, AI companies are recognizing their limitations and are seeking ways to diversify their software offerings to create innovative and engaging user experiences. Recently, Google garnered acclaim in tech circles for its entertaining AI podcasts, with even CEO Sam Altman praising the tool.
With vast amounts of funding still streaming into AI startups in Silicon Valley, consumers can anticipate a steady release of innovative structural experiments that enhance existing tools, such as AI podcast hosts and AI document editors, over the next year. The competition among chatbots is far from finished, and future developments in technology are expected to evolve beyond the conventional chat interface, leaning towards a more complex approach.