Beyond the Chat BoxEssay 10 / 10
Human capability and learning
ESSAY 10

The Best AI Interface Makes the Human More Capable

The measure of good assistance is not only task completion, but what the person can do next.

An assistant can make a task faster while making the person less able to do it independently next time. It can write code, solve a problem, prepare an argument, or make a choice. The immediate result may be good, but the interaction may not help the person understand what happened. A useful measure of AI assistance is therefore not only what the system completes. It is what the person can do after the interaction.

From answer delivery to capability building

A student learning algorithms may ask for a solution. The assistant could provide a complete implementation immediately. Or it could ask what the student has tried, identify a misconception, and offer a hint that preserves the opportunity to reason. Neither approach is always correct. Under time pressure, direct completion may be appropriate. During deliberate practice, a hint may be better.

Let the user choose the mode

A capability-oriented assistant could offer “Do it for me,” “Work with me,” “Teach me,” “Review my attempt,” and “Let me try.” The modes should be easy to change. Someone may want direct help in one moment and a challenge in the next. The system should not manufacture friction or withhold answers merely to increase engagement.

Evaluate beyond task completion

Evaluation might measure later independent performance, confidence calibration, retention, error detection, and transfer to a new problem. A person who finishes one task quickly but cannot recognise a mistake next time may not have benefited in the long term. The goal is not to force learning into every interaction; it is to make learning available when wanted.

Where this can fail

A system may assume every user wants to learn, or that every user wants the shortest path to completion. It may overestimate understanding because the user accepted an answer. Assistance modes should be user-controlled and evaluated with measures beyond clicks, time, or satisfaction.

Questions for research

  1. Which assistance modes improve independent performance over time?
  2. How can interfaces distinguish a request for completion from a request for learning without over-inference?
  3. When do explanations help, and when do they overwhelm?
  4. How can systems measure capability without turning learning into surveillance?

Design hypothesis

The strongest AI interface is not the one that makes the human unnecessary. It expands what the human can understand, decide, and accomplish.

Research starting points

  1. Shneiderman (2020), Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy, International Journal of Human–Computer Interaction.
  2. Amershi et al. (2019), CHI. https://doi.org/10.1145/3290605.3300233
  3. Google People + AI Guidebook. https://pair.withgoogle.com/guidebook/
← Back to all ten essays