Member-only story Chat is dead: Why is it still the #1 designed AI pattern? Why the most basic AI interaction is still most popular, and what it takes to re-invent this Elaine Follow 7 min read · Mar 10, 2025 180 2 Press enter or click to view image in full size by Author Preface: Who are the best at AI prompting? The managers who everyone love working with. Professors who are good at giving instructional guidance. Brand designers who know exactly what visual concept imagery they have in mind. Music artists who know the style, genre, tempo, and fundamental mechanics of making music very well they could teach someone else. Domain experts. These personas give AI enough context. They are hyper-specific in their asks, or appropriately general. They have patience with people, but are impatient with getting work done, which is why they use AI. Chat is the currency of communication with AI systems. Just like how we tell people what to do like an assistant, or riff ideas with them like a thought partner, or co-create with them like a colleague, we write a prompt in a chat. In any of these three major personifications of AI, we talk with AI. We tell AI what to do. We show it images, files, videos, webpages. Sometimes, we show and tell by pointing to parts we care about in these attachments for AI to interpret. “Chat is the command line interface for AI. Can we do better? What is the AI interface?” This question is really asking what for communication methods that are superior to language, or can supplement natural language. Only then, do we graduate from chat interfaces. But in order for a new paradigm to take out chat, this paradigm needs to work as well as chat, or even better than chat, for chat to truly be dead. So chat is not dead yet. But it can slowly be replaced by something more intuitive. AI is sometimes invisible. Recommendation systems, step counters, traffic predictions, autonomous driving, face identification, fraud recognition, spam email identification, inappropriate content flagging. These are all types of AI that run invisibly in the background. There is no chat interface. We are not telling AI what to do explicitly. But we do tell AI implicitly what to do based on the actions we take. More on feedback types here. For example: I like clothes with fitted silhouettes from a handful of designer brands. My search history reflects this. An AI system could make those recommendations accordingly. Press enter or click to view image in full size Revolve I also like to stay on familiar roads when driving, even when a new recommended one could save time. AI can learn this when I consistently ignore rerouting suggestions. I did not tell AI through chat to “stay on my usual roads, and avoid new ones,” but my actions already say it better. Press enter or click to view image in full size Apple Maps Actions are also flexible. Maybe one day I want to try something new. But if AI continues to follow my initial ask to stay on familiar roads only, it will not effectively adapt to me. Intent based outcome specification is not chat, or should not be. It’s instead what we say and don’t say, and exhibit through behavior patterns and actions. People are unpredictable and uniquely nuanced, so sometimes our behavior cannot be explained. This is when chat breaks down. Secret: people are bad at prompting. This is generally well known in the AI world. It is also implied by the many AI educational guides, resources, educational videos and prompt libraries, garbage in garbage out, etc. materials intended to teach people how to prompt better. All this upfront work should help people become more familiar with AI, but prompting still usually take a few tries. Press enter or click to view image in full size Chat interfaces on Mobbin Someone who is really good at prompting will get the exact AI outcome they are looking for in their first try, every time. This is very unlikely today, but gives sense of the ideal bar for human-AI communication. Compared with this standard, we likely all fail. The fewer tries you take, the better you are at talking with AI, means ultimately the more time you save for yourself. To overtake chat takes knowing why this pattern already works well. Chat as an interface is actually an overall smart choice. It allows for mistakes by situating your interaction with AI inside a conversation. Chat implies a continuous back and forth communication between you and the AI system. Chat is fast. Chat is forgiving. Chat maintains a history log. Chat is the continuous sharing and exchange of your stream of consciousness with AI. Press enter or click to view image in full size ChatGPT Of course, it is always hard to describe what you want and exactly how you want it in the first try, just like when talking to someone new. Compare chat with a toggle. A toggle is an extremely reliable UI component. Like a light switch, it is either on or off. It works immediately and is obvious visually. If your intent is to switch it on, one click will immediately do as you wish. Press enter or click to view image in full size Framer components inspired by Dieter Rams’ design principles Ok, now again consider chatting with AI. The problem and advantage for users and AI development teams is that it is literally an empty box — people can ask for anything. Press a button to generate an image, and the only thing you know for certain is an image will be generated. Nobody knows how the image will look or fall among your standards. Press enter or click to view image in full size Capcut desktop video editor How can AI capture the user’s intent as accurately as a toggle? Is that possible, given one is binary and the is not? More importantly, is that desirable? One action takes a millisecond to get the job done, and the other can take who-knows-how-long depending the quality of your prompt and the AI’s ability to interpret your prompt. AI is inherently uncertain. Between actions people can take and know for certain the result, and actions people take not knowing what will happen, is a space for indeterminism from AI. Right now, chat allows people to slowly iterate from version 0 to version 25, or step 1 to step 25, to finally get the outcome. Iterating is fine as a semi-creative process, but I think people also like to save time. Press enter or click to view image in full size v0 by Vercel Saving people time is one core value of UX design. It is truly one of our most valuable assets. If fewer iterations get the job done, then we need to really study AI prompting, or simultaneously design new human-computer interactions that allow for better expression & capturing of user intent. AI models are getting better at interpreting user intent. But UX needs to give AI systems better ways to capture user intent. Words can only do so much. Audio is a step up from words. A picture is worth a thousand words, and videos are worth a thousand pictures. AI isn’t an expert. Pretend you are giving a technical presentation about engineering, design systems, analytics or some specialized topic. Then invite an audience of regular people who are non-experts. In the end, everyone will likely latch on to different takeaways, having heard a different presentation. Now gather this same audience to watch the hottest movie in theaters right now. In the end, everyone will walk away having been captivated by exact same story. The movie was highly visual, with beautiful soundtracks, effective audio, easy-to-follow language, and unbelievably emotional. The movie kept everyone locked in throughout the story. Spoiler: AI is the audience of regular people. Your intent to tell an informative story and for your audience to receive that story through the presentation versus movie was night and day. The presentation is like communicating with chat, but the movie is communicating with everything. As we see, chat is accessible, forgiving, and inherently iterative. But using chat to capture user intent is only from one-dimension. If you want to fully capture the user’s intent from a chat interface, good luck. This is the moment when chat breaks down. The guiding questions become these: Can AI become better at extracting intent from people? Can people gradually become experts interacting with AI? Models are getting better. People are getting more familiar with AI. But the human-AI interaction layer is very much to be designed and very much in need of great UX. It will be a gradual evolution away from chat. Let’s keep moving. Thanks for reading! 👋 For more thought pieces, I write about design & AI Follow me / LinkedIn Why Do AI Projects Fail? 85% AI projects fail, 6 reasons why medium.com The most dangerous thing designers can do In the age of AI, design goes way beyond craft uxdesign.cc