Most talk about AI interfaces starts and ends with the chat box. We wanted to see the whole space we design in, so we catalogued the ways people actually meet AI today and pinned them to one board: 200 interfaces, each with a small sketch of the interaction, real examples and a source.
Open the board. Search it, filter it by family or by how real each idea is, rearrange it by who leads, or read it as an index.
What counts as an interface
Each tile is a distinct way a person acts toward an AI, perceives it, or is acted on by it. Two products with the same interaction count once; the board only splits them when the person's experience differs. Every tile carries two ratings:
- Who leads. From you write, where the AI suggests or responds, to it runs, where it acts unprompted or unseen and you meet the results.
- How real. Mainstream, shipping, emerging, research or speculative.
The tiles fall into fourteen families, from conversation, voice and relationships, through agents, making and robots, to institutions and the invisible systems that rank, price and filter.
What stood out
Typed conversation is one family of fourteen. The chat window is on the board, but most of the space is elsewhere: in the tools people already use, in voices and cameras, in agents working unattended, and in machines that move.
Who leads has two peaks. 58 tiles leave every move with the person and 51 hand it entirely to the machine. The middle is thinner, but it is where much of the new interaction design is happening: agents that ask clarifying questions before they start, plans you approve, permission prompts, steering mid-task, rewinding to a checkpoint.
| Who leads | Tiles |
|---|---|
| You write | 58 |
| You direct | 37 |
| You collaborate | 25 |
| You delegate | 29 |
| It runs | 51 |
Many people meet AI as a subject rather than a user. 31 tiles are institutions and invisible systems: résumé screening, risk scores, insurance coverage decisions, exam proctoring, personalised prices, content moderation. Often the only interface is the outcome.
Robots are learning through new channels. Most of the 16 robot tiles are about how people teach and correct machines: puppeteering a twin arm, wearing a sensor glove while doing chores to make training data, correcting a robot out loud, drawing its path on a camera view, or a remote operator taking over when it gets stuck. FarmBot's web app already writes robot scripts from plain language.
Research is turning model internals into controls. In the thinking spaces family: sliders that turn a concept up or down, panels that show what a model has inferred about you, interfaces generated for a single request, and AI that argues back or withholds its answer until you have decided.
How we made it
We compiled the board in early October 2026 with AI research agents, one each for conversation and voice, software and operating systems, agents and builder tools, bodies and machines, institutions and culture, and HCI research. Between them they gathered 321 candidates and checked the examples against product pages, papers and news reports. We merged near-duplicates and cut the list to 200. Each researched tile links one source; the families and both ratings are our own judgements. A few years are approximate, and some of these products are changing quickly.