There are many ways to present AI interfaces from conversational (ChatGPT) to declarative and closed (Slack summaries) to a hybrid format (Adobe Firefly). At Slack, even though we offer one of the most conversational interfaces in the industry, we debate whether conversational interfaces are currently the most intuitive way for people to interact with AI.
ChatGPT prompt page.
ChatGPT features an open-ended and purely conversational interface.
Slack screen with navigation on the left and threads of conversation on the right.
Slack’s AI summary feature is entirely declarative and closed. Users need only to click a button to summarize the channel. All prompting is pre-prepared to provide the optimal output.
Adobe Firefly screen with 4 AI-generated images and a style panel on the right.
Adobe Firefly features a hybrid interface. It requires an open-ended text input and provides deterministic controls that help fine-tune the prompt.
Tip 3: Make AI design obvious
Open-ended prompts don’t always make for great inputs, especially afghanistan phone number list when users don’t know what your AI can or can’t do for them. So, make it obvious. As designers, it’s our responsibility to ensure that the product is designed in a way that is as intuitive and approachable as possible. In this example, Tome starts with an open-ended text box with premade prompts. As you start to engage, they continue the conversation within the text box. Tome makes it obvious what you can do with the AI and they provide clear boundaries that help people understand what they can’t do with the AI.
Gif of Tome text box showing a list of prompts upon clicking in the text box.
Tome provides suggested prompts that make it clear what the AI can do for the user.
Tip 4: Create boundaries in the UI design
Especially in open-ended conversational AI systems, it’s important to consider boundaries for the types of questions humans can ask and the responses the AI can return. Remember the stories about people hacking a chatbot to make it spit out bizarre and occasionally scary responses? This happened because there was a mismatch between what people expected the AI to do – provide perfectly intelligent and reasonable output – with its limitations as an entirely probabilistic technology that’s figuring out the best tokens to string in a given order.