The Interface Is Not Neutral

Watch anyone use a system they did not choose. They use it for exactly what they are obliged to use it for, and the moment it stops fitting how they actually work, they go around it. A side document. A copy-paste into a notes app. A spreadsheet shadowing the real system, more current than the real system, that nobody has ever mentioned to IT.

I have experienced this regularly, and I have stopped treating it as user error. The workaround is data, and usually the most honest report you will ever get on the gap between how a system was designed and how the work is really done.

I have been reading three papers from the human-computer interaction literature this week, and together they have sharpened something. When we argue about the future of AI, we are usually offered a choice: build models that adapt to how differently people work, or wait for one optimised system to eventually serve everyone. Watching how reliably people route around rigid design, I no longer think that is a real choice. Build AI without studying the workflows and workarounds of the people using it, and what you get is not efficiency but blindness.

Very few are studying what using these things is like

Billions are going into training agents that can drive a computer: click links, fill forms, type on our behalf. Almost nothing is going into studying what it is like to work with one. Understanding User Experiences of Computer Use Agents is largely an argument that this gap exists and that it matters, which tells you how early we still are.

The assumption underneath the gap is that capability is everything. Make the model smarter and the experience improves as a consequence. I have never seen that hold anywhere else in technology, and I do not think it holds here. A highly capable model behind a frustrating interface is still a failure, and it fails in the most expensive way possible, by being nearly good enough to keep using.

The objection I hear is that experience research is premature, that we should wait until the models settle. But nobody is waiting. These tools are shipping to real people now. Deciding not to study how they are actually used does not keep you neutral about design; it just makes it a design decision with nobody's name on it.

The interface is already shaping the user

We talk as though we are in control of our software. Mostly the software is quietly setting the terms.

Language as a Material Interface for Creative LLM Interaction put creative writers in front of a language model through a deliberately constrained physical device and followed them for a fortnight. The constraints did not just make the work slower or harder. They changed the language the writers used and how they structured their time. What a person "naturally" wants to do and what the interface permits are not separable things, which makes the phrase "user preference" far less stable than we usually treat it.

Stop Writing for Me: Generative Refusal in AI Tools for Thought comes at it from the opposite direction. Instead of producing a finished answer on demand, the system withholds and asks the user a question back. People asked more of their own questions, engaged more deeply, and behaved like participants rather than recipients. One design decision, and the cognitive work moved back across the table.

Easier is not the same as better

None of this is new, which is the frustrating part. Healthcare learned it and then forgot it.

Clinical and administrative staff working in older terminal systems became genuinely fast. Keyboard shortcuts, tab stops, muscle memory, eyes never leaving the patient or the page. Then came the move to graphical, mouse-driven interfaces, which look cleaner in a demonstration and test better with people who have never used either.

In practice the fast users got slower, and the response was not to redesign. It was to absorb the loss, or to work around it. This was a key consideration in the replacement of the PAS at The Christie. It should not slow the end user down and impact both patient and administrators.

Do a user's habits shape the interface, or does the interface train the user? Both, continuously, and that is exactly why "intuitive" is such a poor design target on its own. It optimises for the first ten minutes of a relationship that will last years.

The friction question nobody has answered

Mainstream AI tools are tuned for one thing: fast answers and cognitive offloading. Type the prompt, receive the work. For speed, it is superb.

It fails anyone whose job is to think. When the only mode on offer is instant completion, people needing depth end up manufacturing their own friction, slowing the tool down or reassembling their reasoning afterwards by hand. That is a workaround, and by now we know what workarounds mean.

We do not know whether people will tolerate a tool that deliberately makes them think. Friction that serves you still feels like friction, and the frictionless alternative is one tab away. That uncertainty is a reason to study which design choices earn long-term engagement from which people. It is not a reason to keep building one optimised path and calling the people who fall off it edge cases.

What I would tell a client

Stop treating user research as the polish you apply once the real work is done.

Your interface decides what your people are able to accomplish, and over time it decides how they think about the work at all. You cannot serve a diverse organisation through a single copy-and-paste shaped hole. The only route I know is to study how your specific design choices change real behaviour, and to treat every workaround you find as a finding rather than a fault.

User research is not a luxury. It is the foundation.

Sources

• Understanding User Experiences of Computer Use Agents: Design Space and Opportunities for Building Agent UX Prototypes, arxiv-cs-hc, 2026-07-29
• Language as a Material Interface for Creative LLM Interaction, arxiv-cs-hc, 2026-07-29
• Stop Writing for Me: Generative Refusal in AI Tools for Thought, arxiv-cs-hc, 2026-07-29