Transcript: How Canva Is Building AI Into Design | Danny Wu, Head of AI at Canva

Nataraj: Hello everyone, welcome to Startup Project. Today we have Danny Wu from Canva. Canva, to me, is one of the most interesting products that came out in the last two decades. I'm a super user of Canva, and I think Canva redefined this concept of abstraction layers. Previously, when we had to produce anything creative, Adobe sort of worked, but it worked at a very different abstraction layer than what Canva does. What I think Canva really did, in my perspective as a user, is innovate on the abstraction level — the place where I create. They focused on the output I need to generate rather than making me go through what Adobe or other tools require. And that unlocked a lot more use cases for users like me. So I'm super excited to have this conversation with Danny, who's the Head of AI Products at Canva. We'll talk about Canva's use of AI, what they're working on, how the use of AI is evolving within Canva, and any exciting updates coming up next. With that, Danny, welcome to the show.

Danny: Thank you so much, Nataraj. I'm really excited to be here. Quick intro: I'm Danny Wu, Head of AI Products at Canva. I've actually been here for ten years. I started off as a software engineer, and before joining Canva I was freelancing as both a web developer and a graphic designer. What brought me to Canva back then was really recognizing how big the skill, time, and money barrier to entry was for anyone who wanted to channel and turn their creativity into outputs — whether it's a banner, a poster, or a presentation. That's what got me excited about the journey. I do consider myself a creative as well as a technologist, so I'm super happy to talk about the intersection of AI and creativity too. In my day-to-day role, I work with teams and groups all across Canva to focus on how we can bring AI into seamless workflows for Canva users, so we can truly empower the world to design anything.

Nataraj: Talk a little bit more about your time at Canva. You started as a software engineer, then moved into product management, and now you're leading AI products. How did that transition happen?

Danny: Yeah, absolutely. I first joined about ten years ago, back in 2016, which is kind of crazy to remember. Back then we had about fifty people at Canva and maybe one or two million accounts — not monthly active users, just total sign-ups. Barely a few people knew about Canva; we were mostly spreading through word of mouth. As a software engineer, I was building various features, everything from the editor to the brand kit to our first subscription offering, which was Canva for Work. Over the years I led product in areas like content discovery, machine learning, and data science. We've actually been doing ML for quite a long time — since 2018 — for things like ranking templates when you search, graphic recommendations, personalization, onboarding, and lifecycle emails. And of course we got into more non-generative editing AI with tools like Background Remover and other AI-powered editing tools.

The transition, on the surface, felt kind of gradual for us, because before LLMs there had already been a lot of groundbreaking research into generative models like GANs. I'm not sure if you remember StyleGAN, or "This Person Does Not Exist," which was around the 2018–2019 era. Those developments showed that if you feed AI enough data, you can get it to generate things that are genuinely difficult to tell apart from what's real, or from ground truth. In the design and visual space, that immediately had a lot of appeal for us. We had a vision for a world where whatever you wanted, we could generate for you. So if you wanted a really particular decoration or a really particular flower that matches your design — even if we didn't already have it pre-made in our library, and you might not have the budget, the time, or the skills to make it yourself — we could generate it. That's been our thesis for quite a while. And then of course diffusion models, large language models, and transformers hit, and the power of this technology accelerated incredibly. So I transitioned, we started figuring out how to best harness and leverage this technology, and that's what got me to the Head of AI Products role a few years ago.

Nataraj: There's been this transition from focused machine learning models — in 2016 to 2018, recommendation systems were the pinnacle of what we were using in big tech, from the Facebook feed to your Netflix feed. And then suddenly large language models took over the public consciousness in the last three years. Within Canva, how much of AI today is those specific ML models versus the new large language models?

Danny: I'd say the vast majority of our AI is focused around transformer models. When I say transformer models, I'm referring to both large language models and diffusion models and other models for generating images — things like Flux, GPT Image, Nano Banana, and open models. They're all based around the transformer, and that's really the majority of our focus today. There's definitely been a sharp transition. It used to be that for pretty much everything — whether it's recommendations or even something like translation — you would choose or train a model specifically for that task. For different tasks you'd build a different dataset and train a different model with a slightly different architecture. Honestly, I hadn't thought about this too much until right now as we're chatting about it, but it is a little stunning how much attention has shifted over recent years around these technologies, even though ML has been a very strong area with so much value for literally decades.

Nataraj: One of the really interesting things about Canva that I appreciate as a user is this change in the abstraction model. When Canva came about, you started thinking about the outcome you need to generate. You didn't give me a blank canvas like Adobe does, where I decide the layout and go through a zero-to-one getting-started process. Learning Adobe or another visual editor was a lot. You removed that complexity, and we started working at a slightly higher level of abstraction. That, I think, is the biggest innovation I think about with Canva. And now with LLMs, chat has become this abstraction model. My personal thinking is that chat will not work for all creative tasks. So talk a little bit about using chat versus Canva being this place where you execute creativity. As a creative yourself, talk to me about how these form factors are evolving.

Danny: Yeah, absolutely. I actually want to start with your first point about going higher level in the conceptual space, because I think it's very on point. If I look at Canva and think about what it really was, a huge part of it is that we transitioned the world of design from editing pixels to a higher level, which is objects. Instead of using tools like Photoshop or GIMP and having to learn how to cut out a graphic, everything in Canva is designed around manipulating, arranging, or modifying objects — making it simpler to grasp. You can just grab something from our templates library, our elements library, and so on. That's the transition from the pixel space to the object space.

With AI, the way we see it is that design is going even higher level once again, and this time we're going from the object level to the conceptual level. So instead of merely manipulating elements, fonts, and text, you can now conceptualize, manipulate, edit, and create whole ideas, whole themes, and even goals for what you're trying to achieve. You don't have to be so focused on the raster-pixel layer — you spend a lot less time there, the way you did once object-based editing came along. Now you can let AI do more and more object-based editing tasks and focus on an even higher level of abstraction.

On the chat point — fundamentally, when you're creating something visual or graphic, if you have to compress that into just a sentence, or even a longer prompt of a few sentences, maybe with a reference image, that is still an incredibly lossy and restrictive way of communicating with a system. That's why we still very much believe the future of design isn't a world where you just type in prompts, get outputs, and that's the be-all and end-all. At the end of the day, design, graphic creation, and visual communication benefit so much from AI accelerating things, making them easier, and helping you scale — but the most meaningful and most material outputs are the ones that are steered, crafted, or co-made with real human input and human designers. That's why we've been building our design AI, including our Canva foundation model, in a way that enables both AI editing — through agentic design agents — as well as hands-on human editing, so you can just as easily move things around, continue to edit, refine yourself, or change anything.

Nataraj: Can you talk a little bit about some of the AI products you've launched recently, and also this concept of agentic design?

Danny: Yeah, sure. At Canva Create this year we announced Canva AI 2.0 and launched it to our first set of users. It's definitely one of our biggest and most transformative AI launches so far. There's a lot that went on behind the scenes, but what it really is is a platform re-architecture to shift Canva from being primarily a design platform with AI features to a platform powered by an AI orchestrator right in the middle — with availability of all the design tools, all the visual communication tools, and all of our own foundational models. We're still rolling it out and learning, as well as working on the next versions. There were a lot more launches from Canva Create, like Canva Code 2.0. We launched Canva Code many years ago, where you could type in a prompt and get a simple web app or web artifact. With Canva Code 2.0, the difference is that now it's not just an iframe anymore, not just a loose individual artifact — it's actually a page in your design that you can edit with your cursor, with your touchpad, and with Canva AI, of course.

Nataraj: Do you have any thoughts around which use cases AI works better for versus which it doesn't? The reason I ask is there's this whole idea that we'll use AI to make films — that AI will eliminate Hollywood, or visual creativity in general. But whenever I think about it and try to generate a short video or short film, the problem you encounter is that what you visualize in your mind, and how many times you have to re-prompt to get consistency, actually takes a lot of time. If I have to do the same thing — say, realize something I know I can do with Canva objects — it's much faster than me prompting Midjourney or something to get the exact image. That last five percent is where the creative difference between two strong creative individuals actually comes into the picture. So do you think about which use cases AI fits, given Canva's thousands of possible use cases?

Danny: Yeah, absolutely. We think about this in quite a variety of ways because our community's use cases and jobs to be done are incredibly diverse. For example, roughly one third of our user base are either students or teachers, and Canva is used heavily in education — not necessarily just for design or presentations, but also by some schools and districts to deliver course material, like a learning management platform. Then we have enterprise, nonprofits, and of course hundreds of millions of individual users. So we have really diverse use cases.

The way we see the gap you mentioned — between what's possible with really careful, effective, and significant time spent with certain AI technology, versus whether it's actually accessible — is that it doesn't exist equally across all modalities. In the image space it's a bit more clear. If you just want a simple poster, ideas for a logo, a banner, or a graphic, you can throw a prompt into an AI system that has key things like memory, so it remembers anything you said in previous interactions — previous chats and conversations, or previous generations — and knows more about what you're trying to create, your brand, your business, and so on. That's a big part of Canva AI 2.0: things like memory, connectors, and MCP all came with the 2.0 architecture.

Going back to the broader question, the wrong way to think about this is to treat AI — whether in general, or diffusion models, or really big LLMs — as one hammer, and then use that hammer for absolutely everything. It can be quite silly and dangerous to do that. Your video example is a really good one. You can get really stunning outputs from the latest video generation models, but the amount of prompting work and re-iteration, and the time you have to spend — often many minutes at a time before you get a result — means it can very easily take you longer to use AI to create something than it would to create it from scratch or by hand from a template.

So one of the things we're always doing is looking at how different modalities and technologies are developing. A pretty good example is that we recently launched Magic Layers, which has been one of our most popular and fastest-growing features. What Magic Layers does is turn any flat image into an editable Canva design. It extracts not just text from your images — which could be real images or AI-generated, it doesn't really matter — but also all the objects. Not just foreground or background; the model does its best to differentiate all the different objects, generate backgrounds and infills for them, so that you can then move and position everything around. That enables a whole new workflow. Say you're trying to make a poster or a banner: you can use AI to get a few versions, and maybe you really like one, but there are some subtle differences in text, or you want to change the graphics, or make it more on-brand because you want a consistent brand identity. Having the ability to go from AI to AI-made but still fully editable — just like any other design — has been really popular. It wasn't a surprise to us. The average Canva design is edited dozens of times before it's published, and sometimes hundreds of times for things like presentations. So we always knew that for something serious — a material presentation, a sales page, an ad, or even a social media post — there's a degree of refining and editing that you want. The question is more around how much there is, and how do we make it as easy as possible.

Nataraj: Can you talk a little bit about the model evolution? At one point Midjourney was the most powerful model for images, then Nano Banana came, and now there are open-source models from China that are more popular. I also want you to talk about which models you're using, and when you compare implementing these features versus a traditional ML model, how much is the cost difference?

Danny: Yeah, absolutely. I'll start with our general approach, focusing mostly on generative media models. Just a quick recap: there were small toy models maybe five, six, seven years ago or longer — completely different architectures from diffusion — that could only generate relatively small, thumbnail-sized images. Then it was really DALL·E and Stable Diffusion that showed, wow, you can get actually amazing results sometimes. But early diffusion models had all the classic problems — text rendering difficulties, problems with fingers, limbs, and feet — all the things we remember. Those problems have more or less been solved by the latest generation of diffusion models.

One area that's still not perfect is intra-image consistency. If you want a border for a poster and you use any of the latest diffusion models to generate an image, and you look at the borders closely, you'll probably see — just because of how they work, starting from random noise and denoising into an image patch by patch — that it's not perfectly matching or perfectly straight. That probably doesn't matter for a poster, but it does have real limitations if you're trying to do brand-kit transfer and want a poster that follows your branding precisely, uses your logo, follows your branding colors, and uses the right hex color of your logo, not something that's a few percent off. Those are still areas where even the latest-gen diffusion models struggle, but everything is evolving all the time, so I don't expect that to be a problem forever.

Both we and the world have been a little blessed with how quickly diffusion models have advanced. And unlike LLMs — where there's a lot of talk about the rising cost of agentic AI — thankfully, with diffusion models, different patterns, architectures, and use cases mean the technology and the quality are more accessible at every price point than ever.

Nataraj: It's also interesting that when the models inherently get more powerful, it almost creates a roadmap for Canva automatically. The feature ideas are now created by whatever new capabilities the models come out with. Something that wasn't possible to execute as a feature six months back might suddenly become a reality, right?

Danny: Yeah, it very much does. I'd say it's probably about fifty-fifty. Sometimes we recognize something and think, okay, the current models aren't good enough or reliable enough — they just can't do X or Y yet — but we dream of it and want it for a feature, an experience, or a goal. So we put it away in a corner, and when models — either our own or external models — catch up and can do it, we dust off the idea again and build it. And there are also many times when new developments make us realize, wow, that unlocked so many new things; let's sit down and look at everything we can do with it. It's definitely exciting to have new possibilities unlocked. One area that's getting a lot of talk, but honestly not as much attention as it deserves, is world models and how fast that space is advancing. I think it's a bit of a sleeper.

Nataraj: Can you talk a little bit about agentic AI, and if and how you're thinking about using it — what does it really mean to use agentic AI within Canva?

Danny: Yeah, for sure. Agentic AI, to us, is really about the fact that when you come to Canva, you're here to do something. You're here to make a design, edit a design, share something, or maybe you have an output in mind — a goal or job to be done. With agentic AI and Canva AI, we're focused on being the most helpful design partner we can with the technology we have. You can use our agentic AI for everything from routine, tedious tasks — say you have a 20-page presentation and you want to convert the format, or rewrite a disclaimer, or change something referenced across all the pages, and you don't want to go through the effort of finding it on every page and repositioning it so everything still fits. You can just ask Canva: "Hey, can you update all the pages of my presentation? I've got a new logo, make sure it fits." And it's able to do that for you over a few minutes.

That's a fairly simple example. Where it gets really exciting and powerful is when designs become just one of the inputs. We run our own work on Canva, of course — all our docs are Canva docs, all our presentations are Canva presentations. So one of the more powerful things with agentic AI is that we're starting to use Canva more and more to help us inform and make decisions. When we have business metrics or stats we need to pull up, or we want to see why we made a decision and what we considered at the time, Canva has search, it can find and retrieve things from documents, and because it's agentic it has thinking capabilities and can give you recommendations, summaries, and analysis. All of that can come back into a doc that you can then use, edit, and build on top of.

The really transformative value of agentic AI lies not so much in just getting Canva AI to help you design, but in Canva AI being a collaborator that has design superpowers and all your context — connected to your Slack, your Gmail, your Drive, and all the MCPs you want.

Nataraj: One of the interesting things you've also done is the ChatGPT integration of Canva. Can you talk a little bit about how that has helped Canva's business? Did it acquire more customers, get you in front of new customers, or give users a new way to create assets with Canva? How is the impact of something like Canva within ChatGPT being translated into Canva?

Danny: Yeah. Just for a brief intro: with Canva in ChatGPT, Claude, and Copilot, if you install the Canva app, you can get it to create a presentation using a brand kit. Maybe you've done a deep research report or you're already in an AI chat and you want to turn everything into an infographic — you can ask Canva to do that as well. We're available across all the major AI assistants. We first started working with OpenAI; we were one of the first three ChatGPT apps launched. And we've continued to invest, expand, and make it as open as possible. It's not gated to just OpenAI — the Canva MCP server isn't locked down. If you're an AI developer or a startup and you want to talk to the Canva MCP server, just search for it and you can build on top of it.

In terms of value: firstly, if you add up Gemini, ChatGPT, and Claude, there are more than a billion AI-assistant active users. It's almost becoming a pseudo-browser or pseudo-operating system, depending on which metaphor you like. We're always interested in making our design experience as easy and streamlined as possible for everyone. The value of the integration is that we know a lot of our users use Canva alongside AI assistants, and different tools serve different purposes — I use ChatGPT and Codex for some things, Claude for others, and Canva and Canva AI for others as well. We're just trying to make the connections as seamless as possible.

The second part is acquisition and reactivation value. While I can't share specific numbers, we've seen really positive results in terms of growth and net-new, incremental MAUs that these integrations bring onto Canva. When you're using a Canva MCP or app, you're still using Canva to sign in, so we still own and manage the relationship.

Nataraj: One of my personal frustrations with AI-generated text is that I can instantly identify it's generated from Claude or ChatGPT because they have a specific design they output in. A good feature would be: if I'm a Canva Pro subscriber and I attach my Canva Pro subscription to Claude or ChatGPT, I could decide that whenever I say "generate me a presentation," it uses my template within Canva. That would make everything generated more personalized and more reflective of my taste, versus spending another 30 minutes converting what Claude generated into something reflective of my taste. That's been one of the things missing.

Danny: No, 100 percent. I honestly couldn't agree more. One of the interesting things is that this "AI aesthetic" applies to both presentations and images — you can recognize them, and they use similar patterns. It's really interesting because this isn't an inherent downside or limitation of either large language models or diffusion models. It comes as a result of the RLHF and aesthetic-steering process. All these models go through really huge stages of post-training. For LLMs the aim is generally to make them more reliable in agentic coding and more intelligent in coding and other domains. For diffusion and image generation models, the post-training is generally about creating images that humans would prefer and rate higher in blind tests. But one consequence of doing that at a really large scale is that you're essentially taking all these anecdotal signals, smashing them together, and ending up with the median aesthetic of the raters — whether human raters or, increasingly, AI rating which image is better. So you get this imprint of a default steering across all the creative outputs AI can generate. When you ask for a presentation about solar energy, we're still going to apply that RLHF steering to it, and you get that AI aesthetic.

But it's not an inherent limitation of these models, and it definitely has a lot of downsides when it comes to brand, and to the freedom and flexibility to creatively express yourself in any range of style — creating something that feels more like you and your work than like AI's work. We have a lot of active research in this space. For commercial reasons I can't go too deep into it, but I do think the AI look of generated images and presentations is a kind of design collapse, and hopefully we'll be able to make it go away soon.

Nataraj: Yeah, it's also like with Claude Code — you see every website's header and top section looking exactly the same, slightly different color, same fonts. I've been obsessed with not doing that, feeding it design systems and fonts, having my own personal customization. Because in some sense, when generation is so easy, how you present it and put it forward becomes your differentiator as a person. If you're trying to get ahead in your career, who makes the same point in a better way? When you both have the same tools, that becomes the edge, the taste of whatever you're doing. And I think we need more tools — in MCP format, or generally in VS Code or wherever we're vibe coding — because there's a lack of complete design-system input. Right now the design is always the average, like you said, of the internet from the training data, so it always comes out at some midpoint.

Danny: Yeah, very much. It's seen on both ends — whether you're an individual, a startup, or a big business with an established brand system and heavy global marketing. Being able to go the extra mile, from something purely AI-generated to something you can make perfect and add your own final touches to — or generate exactly according to the rules, styles, and brand systems you have — is really important. Companies spend something like hundreds of billions on marketing and advertising, generally for spreading visual messages. And as every marketer knows, a good creative can easily make a campaign successful, and a bad creative can tank it even if you have the best targeting or the best online product.

One of the sadder things about generation being so easy is that, broadly as a society, I think we still instinctively and intuitively care about quality, and it absolutely matters. I'm not just saying this because I have some particular axe to grind. We see it ourselves: when we use AI for marketing, there's a huge difference between using AI entirely hands-off — just letting an agent design something and seeing if it performs in performance marketing — versus using AI and then taking a few minutes, or ten minutes, of human effort to polish it and make sure it's actually on-brand. The differences in results are honestly really substantial.

Nataraj: It's also about how your eye evolved over millions of years — we're very visual people. Even though people might not consciously care about it, they're subconsciously programmed to care. That's the real reason you still have to do these things. If you want to stand out, separate yourself from the crowd, or if everyone is generating the same thing, the only way to differentiate is to put in this additional effort of bringing your own design system. That's a way to show creativity and taste. You become the tastemaker when all the content search and curation is happening by AI. I think that's where people will find their edge when they're working. Can you talk a little bit about your process inside Canva for picking a new feature and getting it to production? A new model launches — how do you take a new idea to a product a customer might use? What are the different stages you go through?

Danny: Yeah. One of the first things we try to do is settle Canva into a collection of different goal streams that we like to call "little startups." They're not literally individual startups, but the framing is that we try to make things as multi-threaded and parallelized as possible, with different goals and different streams around certain things.

All our AI features are a little different, but generally it often starts with a prototype — someone having an idea, a scratchy feeling in their head, and they write code or test something out. It gets shared around, and people say, "This is actually quite useful," or "Wow, there might be something there." Then we might do a spike, where we build a more robust prototype over a couple of days and see if the idea has legs. If it does, it progresses through what we call the "chaos to clarity" spectrum. At the beginning there are so many unknowns — including how much to invest in the idea, whether to invest at all, and how to do it. As we work on it more, things get clearer. We like to dogfood our own products a lot, so features go through internal testing before a one-person rollout, a ten-person rollout, and so on.

With a lot of model upgrades and changes, we try to isolate the impact of the model versus anything else. So whenever we can avoid it, we don't launch an upgrade to a feature and a new model at the same time. We keep everything about the UI, the interactions, and the patterns the same, and just swap out the underlying model for a different one. This lets us isolate differences more specifically to one source, so we can tell whether a model is actually better for the use case in the harness we're using.

Nataraj: Are there any interesting ways customers are using Canva's AI that you never intended?

Danny: Yeah. One of the more popular, interesting, and surprising ways people use Canva Code is that a lot of teachers use it to create self-grading quizzes. They essentially build quizzes where it's not just completing a quiz — the teacher sets what the correct answers are, and after the student completes it, the scores, names, and completion data are all stored. This isn't something we ever thought Canva Code would specifically be useful for, or that it would be so popular. It just goes to show that with prompt boxes you have so much amazing possibility, and you never really know what our hundreds of millions of users will end up using it for. There are definitely things that surprise us.

The other example I'd give is that we're starting to use Canva AI a lot more as a workplace AI assistant internally — finding information, creating reports, synthesizing things, or even getting advice on how to navigate an ambiguous situation. That's still something we're heavily working on and dogfooding, but it's another thing I'm quite excited about.

Nataraj: Do you see Canva overlapping with Figma over time?

Danny: There's already some overlap today. The way I ultimately see it is that we're focused on our own race and what we can do to deliver value for our community, less so around others. On a personal note, I think Figma is a really great product, and we use Figma for our product design. To us, we're about impact — about making design as successful as possible so that everyone can create just about anything. Our mission is focused on delivering a design platform that caters across all skill levels, professions, and needs, and across anything that's possible to design. Just as we have docs, code, presentations, and video, I also don't see the design market as one that's going to be entirely dominated by one tool or company across the entire space. And I actually don't think AI is going to change that. So I hope that answers your question.

Nataraj: One last question as we near the end of the conversation. Are you planning to launch more video-related products within Canva?

Danny: Absolutely. One of the things we're working on is a very significant rewrite and re-architecture of our video editor. One of the core platform problems is that, back when we built Canva Video for the first time, the average video uploaded and published was 720p or maybe 1080p at first. Phones streaming 4K was a high-end feature when Canva Video first launched. So our stack, since the beginning, wasn't strongly designed for the kinds of content and interactions people — especially on mobile — are working with now. We have an internal effort where we're really investing in revamping our video editor, and that comes down to performance, functionality, AI, reliability, and so much more. It's a very active area of investment for us.

There are a few reasons. Part of it is the increasing amount of content creation and consumption happening with video. And part of it is that we feel there'll be an even bigger disruption when generative video becomes truly accessible, low-cost, fast, and easy to use. We're quite excited for what that can mean for creativity and for empowering our users, so we're investing a lot there.

Nataraj: I think that's a good note to end our conversation. Thank you, Danny. Thank you for coming on the show and sharing all about what you're working on at Canva.

Danny: No problem, thanks so much. I really enjoyed this, and thanks for having me again.