The “Chat” in ChatGPT’s name neatly encapsulates a major element of why OpenAI’s AI-infused assistant has had such epoch-shifting impact. We’ve been using PCs and phones to chat with humans for years—at least since the era of AOL Instant Messenger. Talking to a piece of software, using a largely similar interface, is not that vast a mental leap.

And yet ChatGPT’s chat-centricity is also a bit of a stumbling block. Not every task lends itself to a pure back-and-forth dialog. Writing, for instance, calls for something that looks more like a word processor, with a roomy window to compose text and the ability to make as many edits as you want. That need doesn’t change just because you happen to be collaborating on your writing with an AI assistant.

Last week, OpenAI released a version of ChatGPT that’s mindful of this scenario. Called ChatGPT 4o with Canvas, it’s a new option in the Model menu for ChatGPT Plus subscribers, designed with both writing and coding in mind.

At first blush, it looks just like . . . well, ChatGPT. But if you ask it to write something, it may decide to put the results in a large document window and relocate the chat to the left. If you make further requests, it will update the window rather than starting all over. You can edit the text yourself, or fire up a blank document and plug in material of your own, such as something you’d like to get summarized. There are also options for polishing prose, checking grammar, adjusting reading level, and shortening or lengthening text.

For text-heavy AI jobs, Canvas radically improves ChatGPT’s old interface. At the moment, however, it also feels incomplete. I couldn’t get wide swaths of it to work in Safari on a Mac or iPad, and at one point it inexplicably ate some text I’d previously pasted in. (Fortunately, you can revert to earlier versions.) The feature could also use at least basic tools for wrangling the documents it creates, such as the ability to rename them and see them all in one place.

Conceptually, Canvas brings ChatGPT closer in interface and functionality to a Google AI tool called NotebookLM. A web-based app released late last year, it’s a research tool optimized for writers whose work involves reference materials—text files such as interview transcripts, web pages, PDFs, and even audio files. Feed such items into the app, and it will summarize them, answer your questions about them, and even suggest questions you should be asking.

Much of what you can do with NotebookLM can be approximated with any AI chatbot. Still, I’m impressed by the quality of its summaries and its ability to pinpoint relevant facts more swiftly than I can locate them on my own. Like ChatGPT with Canvas, aspects that don’t relate directly to AI aren’t fully fleshed out: I certainly wish I could reorder the notes I create, for example. Nonetheless, it’s on the short list of AI tools that have had a genuinely work-changing impact on my productivity.

I am, however, more amused than aided by Audio Overview, a new NotebookLM feature that’s currently going viral. Give it a few minutes to crank away, and it will churn out a simulated podcast starring two NPR-doppelgänger robohosts who discuss whatever you’ve uploaded into the app. The human-level fluidity of their chatter is uncanny, though the substance tends to be superficial and their manufactured enthusiasm for whatever they’re talking about can be hilarious. (Just as a test, I created a podcast in which they discuss what it’s like to watch paint dry. Turns out it’s fascinating!)

Audio Overview is a dazzling, weirdly transfixing technical achievement. I’m tempted to spend half my time shoveling stuff into it just to hear what comes out. Yet it feels out of place in an app otherwise devoted to serious research. It’s a little as if a seminar on library science brought in a magician to perform sleight of hand tricks involving books, and is potentially a worrisome sign that NotebookLM is losing track of its purpose.

The bigger risk is that neither NotebookLM nor ChatGPT 4o with Canvas will reach maturity and then stick around. Google (which calls NotebookLM “experimental”) is legendary for abandoning products—even some pretty beloved ones. And OpenAI (which says Canvas is in beta) is so busy trying to boil oceans, from search to AI hardware, that it’s impossible to know where Canvas might fall among its many priorities.

All of which makes me even more grateful for a couple of time-tested apps I find indispensable. After deciding to wean myself off of Evernote last year, I now do most of my note-taking in an elegant Mac/iPad/iPhone app called Bear. For writing projects involving research materials and more than a few hundred words, I use the iPad version of Scrivener, an app with a legendarily devoted following among long-form writers.

Along with being excellent, Bear and Scrivener have another thing in common: They’re the flagship products of tiny companies. Both are thoughtful, focused, and feature-complete in ways that happen most often when a developer has everything riding on building no more than a few products and making sure they’re worth paying for.

What neither Bear nor Scrivener has are any obvious uses of artificial intelligence. Whether that reflects a considered decision on the part of their creators or just a lack of resources, I’m not sure. It bothers me only because the day may come when being unable to compete on AI with OpenAI and Google puts small developers at such a severe disadvantage that their continued existence comes into question.

But wait. Bear and Scrivener are about to get some fairly meaty AI functionality without lifting a finger. Apple’s new Apple Intelligence AI features, which are due to ship this month, include several writing tools that will be available across iOS, iPadOS, and MacOS, including options for rewriting, proofreading, summarizing, and formatting text. While they don’t replicate the most ambitious aspects of Canvas and NotebookLM, they’ll give every app that involves text input a welcome baseline of AI savvy.

Meanwhile, I do hope that ChatGPT with Canvas and NotebookLM transcend their present rough-draft status. Like many products of generative AI, they’re already amazing. But that’s not enough to guarantee they’ll be keepers.