Adobe’s conversational AI agent is a mediocre design intern

AI image tools rarely make me feel like I’m part of the creative process. They are, after all, mostly designed so that people with no design experience can type in a few words and get back a usable result. So I was pleasantly surprised by Adobe’s latest take on an AI image assistant: It’s a bot designed to take away some busywork, while still granting you creative control.

Adobe’s conversational AI agent is a mediocre design intern

The Firefly AI Assistant isn’t as good as a professional human designer or photo editor, but it’s fun to watch it work.

Unlike AI generators that are specifically designed to make and edit images or video, Adobe’s Firefly AI Assistant, which I’ve been testing in beta, is more like a multitasking middleman that can operate Adobe’s design apps for you. On its website, Adobe says that you can just “tell Firefly AI Assistant (beta) what you need, and it will use tools from apps like Photoshop, Illustrator, and more to complete multistep projects in moments.”

The user interface looks like a typical chatbot. There’s a text box you can type prompts into, and a plus symbol for uploading media files. It doesn’t use the actual Adobe apps on your computer, but it has access to common capabilities like masking, object detection, and image generation. The AI assistant is designed to be conversational, so you can ask the chatbot to “make this photo more colorful,” and it’ll do so while explaining its actions.

Photo edits and illustrations completed by Adobe’s AI are convincing at a glance. It changed my hair color in one photo and then the background location and lighting in another. The results aren’t perfect: some had colors that were too vivid or alterations that hadn’t been properly blended into their surrounding environments. But I suspect the average person wouldn’t assume that my results were made or manipulated with AI — it just looks like the work of a novice designer.

What actually makes the Firefly AI Assistant intriguing is how it interacts with you. I gave it a picture of my cat by a window and asked it to make the sky cloudless and sunny. It didn’t just go “sure” and give me the edited image. The chatbot described the scene in the pre-edited image with surprising detail (it correctly identified that my cat is a Maine Coon despite the photo mostly just showing his ass), and then explained how it’s going to achieve the results I’d requested. It mentioned specific tools from Photoshop and Lightroom using established editing terminology, explaining the process step by step. You don’t actually get to see the image being edited in real time, but the chatbot will tell you which features it’s using to achieve each result.

The Firefly AI Assistant is also surprisingly forthcoming about its limitations. When I asked it to separate objects from a JPG file into separate layers, Firefly said it couldn’t do so, but offered two different courses of action for splitting the image into separate elements, explaining the pros and cons of both. After I chose one, the bot then described its editing process, including the fact that what it was doing wasn’t working. “I notice the gaussian blur approach isn’t giving me true transparent cutouts — it’s outputting full-image PNGs,” it wrote. The chatbot redirected itself and used masks and Adobe’s image cropping and resizing tool instead.

You can also ask the chatbot to add new objects into images, akin to Photoshop’s Generative Fill or Google’s Magic Editor features. It didn’t hesitate to add cigars, doobies (erm, “hand-rolled cigarettes”) and even guns to my photographs, but refused to generate anything outright illegal. I could make an obviously fake album cover of myself pointing a gun at the “camera,” but not make myself look like I’m shooting anyone. The results for these sort of edits are also visually subpar compared to asking for things that won’t necessarily require generative AI tools, such as lighting adjustments, but I wouldn’t say they’re outright bad — they’re just not good enough that they could fool me. It also refused to alter the shape or size of my face and body, or put me in revealing clothing — something Grok could use some notes on.

Prompting chatbots usually makes me feel like I’m asking a theme park mascot for directions — the constant enthusiasm is unnerving, and not at all like talking to another adult human being. The Firefly Assistant is still guilty of gushing out pointless and unnecessary praise (I do not need it to tell me that my edit request is a “great idea”), but for the most part, what it’s saying is actually useful.

When the bot needs additional information, you’ll be prompted for it. I asked it to turn a photograph I took of two cocktails into an illustration-style graphic that can be used to advertise a bar on social media. It then asked me what platforms I was planning to post the design on, providing a list of options like TikTok, Instagram, and Facebook that I could select, alongside the standard pixel dimensions for such content. If I wasn’t already familiar with Instagram Square posts being 1080x1080p, I’d have come away from that experience having learned something.

This is the most intriguing way a creative chatbot has interacted with me so far. By asking it to make changes I never bothered to learn how to do myself, the AI assistant shows me where I need to focus on building those skills by explaining its own workings. I’d say it’s less useful than asking Google Search or YouTube for tutorials, but those services won’t complete the task for you while you learn. I personally enjoy manual creative processes too much to ever delegate that to a chatbot in the real world, but for folks who want to make edits that wouldn’t otherwise be worth their time? I can see the appeal for this.

Canva also recently launched its own conversational design agent. That too, is rife with tendencies to communicate through flowery language and infantilizing praise, but it doesn’t explain its working process like Adobe’s chatbot does, and the results didn’t quite hold up to what I saw Adobe’s producing. You just give the Canva assistant instructions and keep prompting until you’re happy with the outcome. For those willing to learn, Adobe’s tool may actually help to demystify some design and editing basics while delivering on your requests.

But Adobe is mostly pitching this Firefly assistant as a means to save creative professionals time by undertaking labor-intensive tasks. I have middling editing skills at best, and I don’t feel the chatbot would be useful to me unless I’m happy to pump out sub-quality work. For people with very established design skills, I fear using this would feel more like babysitting a new intern than having a helpful colleague. It may become easier to pitch this AI assistant to creatives if it can deliver work that’s indistinguishable from something professionally edited. But for now, it’s far too much of a novice.