Learning outcomes
- Communicate capability and uncertainty at the right time
- Support correction undo and recovery
- Calibrate trust without anthropomorphic claims
Mental model
Design an AI interaction as a calibration loop: set an accurate expectation, show consequential system activity, let the user inspect evidence, make correction cheap, and learn from the correction without hiding uncertainty.
Theory
Before interaction, explain what the system can do through available actions and constraints rather than marketing claims. During execution, expose durable states such as queued, retrieving, awaiting approval, running a tool, verifying, failed, and complete. After execution, distinguish generated content from retrieved evidence and tool results.
Corrections should target the smallest meaningful unit: edit an input, remove a source, retry a failed step, undo a side effect, or report a wrong assumption. Preserve the user's original intent and show what changed. Confidence displays should represent calibrated evidence or known system state, not decorative percentages generated by the same model.
Alternatives and trade-offs
Progressive disclosure keeps routine work compact while allowing inspection. Detailed traces help expert users but can overwhelm general users and expose sensitive reasoning or data. Automatic personalization reduces effort but may create stale assumptions; explicit user-controlled preferences are easier to audit.
Failure modes and misconceptions
Typing animations are not operational progress. Anthropomorphic language can inflate trust. A generic retry button may repeat the same failure. Explanations without evidence can sound convincing while remaining unverified. Removing user control in pursuit of automation makes exceptional cases harder to recover.
Knowledge check
Which interface state helps a user distinguish slow useful work from a stalled or unsafe action?
Decision scenario
A research assistant shows search, retrieval, synthesis, and citation-check states. The user can remove a weak source, inspect the affected claims, retry only synthesis, and restore the prior answer without rerunning the entire workflow.
Learning outcomes
- Explain AI Feedback and Correction as a system mechanism rather than a slogan.
- Compare its alternatives, trade-offs, and production failure modes.
- Apply the concept to a decision and identify evidence that would validate it.
Trade-offs
Using AI Feedback and Correction can improve capability or control, but it also introduces cost, latency, complexity, and failure modes that must be measured against an explicit objective.
Evidence assessment
Theory and decision mastery
Decision scenario
A production team must adopt AI Feedback and Correction while meeting quality, latency, security, and operating constraints.
Which decision process is most defensible?
Primary sources
- Guidelines for Human-AI Interaction — Microsoft Research / ACM CHI, verified 2026-07-21