# Invisible Users, Invisible Interfaces: Accelerating Design Iteration with AI Simulation

Alex Liss, Huge | AI Engineer World's Fair 2025 | 12:37

Source: https://www.youtube.com/watch?v=8D_VdU6DBhI
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/invisible-users-invisible-interfaces-accelerating-design-iteration-with-ai
Published: 2025-06-03
Tags: computer-use, design, evals, multimodal

## TL;DR
- AI should help teams discover user needs and test designs, rather than being added as a chatbot feature.
- Intelligent Twins can simulate different audiences completing tasks across many interfaces and markets.
- Computer-use models, human observation, and repeatable evaluation can turn simulation results into focused design briefs.

## Summary
Alex Liss argues that many AI products have created a trust gap by adding unreliable chatbots to existing websites and apps. He returns to Don Norman's idea of invisible interfaces, where software feels simple enough that people forget they are using it. Liss proposes using AI earlier in the design process, during need finding. Huge's data platform, Live, provides demographic, psychographic, and contextual data for audience simulations. These data become Intelligent Twins, which represent user behaviors, motivations, and desired outcomes. Designers can brief them to complete tasks and evaluate interfaces at different speeds and scales. Liss illustrates the method with an audit of sports websites, using casual fans and super fans across navigation, information architecture, and engagement tasks. He describes the approach as experimental and calls for standard briefs, audience dimensions, run counts, and test-control comparisons before judging where simulated users help human teams.

## Key ideas
### The current AI trust gap comes from unreliable features being pushed into familiar products
[00:25](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=25s)
Liss cites Edelman research from December 2024: 32% of US adults said they trusted AI, while 44% of adults globally said they felt comfortable with how businesses were using it. He connects this discomfort to what he calls AI slop, such as search results telling people to eat rocks or a website claiming a car costs $1. Many products add chatbots and label the result as magic, even when the interaction is plainly wrong. Liss says the useful change is that people can now speak to a machine learning model in natural language, which creates a reason to rethink user experience instead of decorating existing software with AI.

### Invisible interfaces give AI design a clear target
[02:00](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=120s)
Liss draws on Don Norman's writing about simple, efficient everyday products and the idea of the invisible computer. An invisible interface feels so natural that users practically forget they are using software. He proposes using AI to help create those experiences by accelerating need finding. He is explicit about his role: he is a data scientist, not a designer, so the talk presents a process rather than finished visual designs. The goal is to understand user needs faster and give design teams better evidence before they decide how an interface should work.

### Simulation can turn collected audience data into active participants in design
[03:09](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=189s)
Liss compares the proposed process with pilot training. Pilots learn to handle complex cockpits through repeated simulation, while traditional design gathers qualitative observations, quantitative data, and ethnographic material before prototyping. His proposal turns those data artifacts into invisible users that can participate in a smaller feedback cycle. The broad process still includes defining an audience, mapping intentions, identifying tasks, analysing the process, refining insights, and creating alternatives. AI changes how some of those steps work by allowing simulated audiences to interact with proposed experiences during need finding.

### Intelligent Twins model user behavior, motivation, and desired outcomes
[05:04](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=304s)
Huge's Live data platform combines demographic, psychographic, and contextual data to represent audiences. Liss turns that foundation into an Intelligent Twin, a simulation of a set of user behaviors and desired outcomes, including needs and motivations. A team can brief these twins to evaluate a specific interface through specific tasks, much as a human designer might receive a brief for a heuristic analysis. This makes the audience model active during the design process. It can respond to a defined interaction rather than remaining a research document that designers consult only before prototyping.

### A sports-site audit shows how simulated users can work across audiences and tasks
[06:39](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=399s)
For a sample project, Liss used two personas: a casual fan who was newer to sport and a lifelong, knowledgeable super fan. Across three sports websites, the simulated users received tasks covering navigation, information architecture, and fan engagement. There were four tasks in each category, producing 72 simulated actions. The audit covered basketball, the Olympics, and the English Premier League. Liss says navigation performed fairly well across the sites, while task completion fell as users moved into content browsing, information architecture, and engagement pathways. The method could therefore compare broad categories while retaining differences between audience types.

### Simulation can move from a global category view to focused design briefs
[07:52](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=472s)
Liss says the method can roll findings up to a high level or examine friction in a specific part of an experience. Its different levels of detail can support design briefs that are broad, narrow, or deep. The closing stage combines computer-use models, computer vision models, and human observation in the loop. The resulting brief points the human team toward pain points that distract users in a category. In his example, the value is not a replacement for design judgment. It is a way to help people spend more time solving the problems that the audit has surfaced.

### Faster implementation increases the need to agree on why a product should exist
[10:06](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=606s)
Liss points to progress in the first half of 2025, including Anthropic's MCP protocol and possible paths from Figma prototype components to code components in React, Node.js, and other frameworks. As tools make implementation easier, he says teams need to focus more on the reason for a design and the user problem it should solve. The technical how becomes easier to produce. That makes the quality of the initial strategy and need finding more important, because a team can now generate a large number of solutions without necessarily choosing a useful one.

### The method needs repeatability and comparison with human teams
[10:53](https://www.youtube.com/watch?v=8D_VdU6DBhI&t=653s)
Liss describes the approach as experimental and early. He wants his company to standardize the method in a code repository, including briefing instructions, simulated audience dimensions, the number of audit runs, and task-completion and failure parameters. He also proposes test-and-control work to isolate where Intelligent Twins help with design need finding. Those comparisons should cover different industries, geographies, and domains, and should examine how simulations work alongside human teams. His conclusion is restrained: better evidence from AI simulation could help teams create clearer interfaces, but the method still needs disciplined evaluation.

## Notable quotes
- "The real magic of GenAI is the fact that users can actually talk to a machine learning model in natural language, which has never been possible before." (01:15)
- "We want to use AI to design interfaces that actually feel like magic, not by stuffing chatbots into the website, but by accelerating need finding." (02:25)
- "Intelligent twins become an active participant in the design simulation process." (05:49)
- "Users don't need more websites that have GenAI chatbots in them that don't work." (12:00)

## Tools & references mentioned
- Edelman
- Cassie Kosak
- Don Norman
- Nielsen Norman Group
- The Invisible Computer
- Huge
- Live
- Intelligent Twins
- Anthropic
- MCP
- Figma
- React
- Node.js

## Who should watch
- Design and research teams that need to test audience assumptions across many markets or user types before committing to interface directions.
- Product teams that are adding AI features to existing software and want to understand why those features create friction.
- Engineers and design-technology leads evaluating computer-use agents or simulated users as part of an existing research process.

## Related talks

- [Form factors for your new AI coworkers](https://aietalks.com/talks/form-factors-for-your-new-ai-coworkers) (Craig Wattrus, Flatfile, 15:35)
- [The Intelligent Interface](https://aietalks.com/talks/the-intelligent-interface) (Samantha Whitmore & Jason Yuan, New Computer, 18:56)
- [Designing AI To Scale Human Thought](https://aietalks.com/talks/designing-ai-to-scale-human-thought) (Jun Yu Tan, Tusk, 12:24)
- [Second Order Effects of AI](https://aietalks.com/talks/second-order-effects-of-ai) (Cheng Lou, 21:46)
- [Good design hasn't changed with AI](https://aietalks.com/talks/good-design-hasnt-changed-with-ai) (John Pham, SF Compute, 20:25)