MSc Research · Kingston University In Progress Recruitment Open

AI that knows
before you ask.

A live MSc research project on OS-level intent inference — designing AI affordances that surface from how you're already working, not from a chat window you have to open.

Participate in the Study → Read the Research View AI Drop →
Computational Intent Pipeline
Prototype Demo Coming Soon
Signal Inference Resolution Affordance User Control
01 The Problem

You have to ask. Every time.

Every interaction with current AI starts with a context switch. Even for tasks where AI would obviously help, the friction of getting there is often higher than the help it provides.

01
Stop the work
02
Open a chat window
03
Formulate a prompt
04
Re-integrate the result
17–41s
saved per task in the prior published study — just by replacing "open browser" with drag-and-drop. This thesis asks what happens when we remove the invocation step entirely.
02 The Idea

From explicit invocation
to passive intent inference.

The Computational Intent Pipeline is a five-layer architecture that reads ordinary OS-level behaviour, infers a likely intent, and surfaces a relevant AI affordance — quietly, in place, dismissible at any moment.

01 / Signal
Signal Layer
Captures OS-level behavioural events via macOS APIs.
02 / Infer
Inference Layer
Classifies signal combinations into intent classes.
03 / Resolve
Resolution Layer
Maps inferred intent to a specific AI task.
04 / Surface
Affordance
Presents a minimal UI element at the relevant point.
05 / Control
User Control
Accept, dismiss, or redirect — every event logged.
No prompt required. The system reads behaviour, not text input.
No context switch. Affordances surface inside the workflow you're already in.
You stay in control. Every suggestion is dismissible without explanation.
Ambient, not interruptive. Designed to be ignorable when the inference is wrong.
03 Three Intent Classes

What the system is listening for.

The MVP narrows the surface to three intent classes — chosen because each maps to a recognisable signal pattern and a well-understood AI task.

Class 01 / Transform

Translation & Transformation

You paste text from a foreign language into a document. The system detects the language or style mismatch and surfaces a translate affordance — before you've opened a new tab.

Signals: paste event lang detect style delta
Class 02 / Comprehend

Comprehension

You've been scrolling slowly through the same dense paragraph for half a minute. The system detects the dwell pattern and offers to explain or summarise — right where you're reading.

Signals: dwell time scroll velocity re-read loop
Class 03 / Cross-Reference

Discovery & Cross-Reference

You've switched between three tabs on the same topic five times in two minutes. The system detects the pattern and offers to synthesise across them — before you open ChatGPT yourself.

Signals: tab switching topic similarity app focus
04 The Prototype

A native macOS build, in development.

The prototype extends AI Drop — a shipped, published macOS application — so the new intent-inference layer can be tested in a realistic, real-app environment rather than a sandbox.

Build Specification

  • Platform Native macOS, Swift
  • Foundation Extends AI Drop (open source)
  • Signal APIs NSPasteboard · CGEventTap · NSWorkspace · Accessibility API
  • Inference Rule-based heuristics (MVP); lightweight ML under evaluation
  • Telemetry In-app behavioural logging — timestamps, intent class, affordance shown, user response
  • LLM Access Bring-your-own-key
In Development

What you'll see here later: short demo clips of each intent class triggering in context, plus screenshots of the affordance UI and the in-app event log. This section will update as the build progresses.

// status build log will populate as milestones land
// next signal capture layer — integration tests
// then first intent class wired end-to-end
// then affordance UI — first pass
05 The Study

A within-subjects evaluation, three conditions.

Participants complete equivalent document tasks under three invocation modes, counterbalanced for order. The aim is to measure whether intent-mediated AI feels faster, lighter, and more trustworthy — or whether it doesn't.

Condition A

ChatGPT

Chat-mediated baseline. Open the window, formulate the prompt, paste back.

Condition B

AI Drop

Triggered invocation via drag-and-drop notch — the previously published interaction.

Experimental Condition C

Intent Inference

The new prototype. Affordances surface from behavioural signals — no explicit invocation.

· What Is Measured
Task Completion Time Via prototype interaction logs.
Cognitive Load NASA-TLX, per condition.
Trust Jian et al. (2000) trust scale.
Qualitative Observation Phase 1 contextual notes.

Analysis Wilcoxon signed-rank tests with effect size r for paired comparisons.

Participant Recruitment · Open

Today's AI waits to be asked.
We're building AI that knows when to help.

Be part of the study that tests whether that's actually possible. A single 45–60 minute session — your patterns shape what comes after chatbots.

What You'll Do
Phase 1 — Observation 20 min
Two everyday document tasks on macOS. No prototype, no instructions — just how you'd normally work, while I observe.
Phase 2 — Three Conditions 25–35 min
Equivalent task sets under ChatGPT, AI Drop, and the new prototype. Short cognitive-load & trust ratings after each.
Who Can Take Part
  • macOS as your primary operating system
  • Regularly work with documents — writing, editing, translation, research
  • Aged 18–65
  • No prior involvement in this research team's studies
What You Contribute

Your real work patterns and reactions provide the first empirical evidence for or against a new way of designing AI into operating systems — one that acts when you need it, not when you ask.

Express Interest → Opens your mail client — pre-filled subject, no obligation.
06 Foundation

This is the direct sequel to AI Drop.

The prior study found that where AI sits in the interaction matters more than how powerful the model is. This one asks what happens when AI doesn't wait for you to reach for it at all.

Prior Publication
Adapting Established GUI Paradigms for Generative AI
Wallbrecher & Kiruthika, 2025
View AI Drop →
n = 30
Within-subjects
17–41s
Time saved / task
p < .001
All conditions, d > 0.8
What that study didn't do: infer intent, work across multiple apps, or surface AI before the user initiated. This thesis does all three.
07 Reflection

The questions this work is actually about.

01 /

The problem with AI is not intelligence — it's interaction distance.

02 /

Reducing friction is not a UX nicety; it determines whether people use powerful tools at all.

03 /

The question is not "can AI infer intent" — the question is at what cost to trust and autonomy.

04 /

This thesis tries to answer whether proactive AI feels helpful or intrusive — and what design decisions determine which way it goes.

Open questions. Findings will appear here as the study concludes.

08 About & Status

Who, where, when.

Project
  • Researcher Moritz Wallbrecher
  • Programme MSc User Experience & HCI
  • Institution Kingston University London
  • Module CI7801 — User Experience Major Project
  • Supervisor Jay Kiruthika
  • Foundation Wallbrecher & Kiruthika (2025)
Status & Contact

This page is a living document. Sections marked In Progress will update as the prototype is built and the study concludes.