Active Research · CU Anschutz DBMI

ClarifAI

Turning vague AI feedback into structured, actionable insights for designers.

The Problem

Vague feedback tells designers nothing.

Automated Online Feedback Gathering (AOFG) tools are widely used to collect responses from users of AI-generated outputs — but the feedback they produce is often too vague or ambiguous to act on. "Was this helpful? Kind of." gives a designer nowhere to go.

ClarifAI intercepts this feedback before it becomes noise. A three-module LLM pipeline filters out irrelevant responses, flags vague or ambiguous comments, and conducts a targeted follow-up dialogue to turn them into structured, actionable insights.

ClarifAI Feedback Dialogue
You rated the AI-generated image 2 out of 5. Can you tell us more about what didn't work?
It looks a bit off — I'm not sure how to describe it.
Thanks — was the issue with the composition, the lighting or colour, or the subject details (e.g. faces, hands)?
The hands — they look distorted and have too many fingers.
Got it. I've logged: anatomical distortion in hands — extra fingers, unnatural proportions. ✓ Saved

How It Works

Four steps to actionable feedback.

1

Collect Feedback

Users rate and comment on AI outputs through a standard AOFG interface embedded in the web application. This is Stage 2 of ClarifAI's three-stage study platform, which also covers prerequisites and consent (Stage 1) and a discussion board (Stage 3).

2

Telemetry — Relevance Filter

The Telemetry module classifies whether each piece of feedback is actually about the AI output. Off-topic or tangential responses are filtered out before they enter the pipeline. In our evaluation, the Telemetry module achieved 100% precision on relevance filtering.

3

Flight — Vagueness & Ambiguity Detector

The Flight module classifies relevant feedback as either specific enough to act on, or vague/ambiguous. Only feedback that fails this check is escalated to the clarification dialogue, keeping the experience lightweight for users who already gave clear responses. Flight achieves 94%+ accuracy on vagueness and ambiguity detection.

4

CapCom — LLM Clarification Dialogue

The CapCom module engages the user in a short, targeted follow-up conversation. An LLM asks only the questions needed to resolve the vagueness or ambiguity, then produces a structured, machine-readable summary — a specific, categorised insight that designers can act on directly.

Why It Matters

Better feedback. Better AI.

Actionability of Feedback by Condition

ClarifAI clarified feedback High
Raw vague feedback (unprocessed) Low–Medium
No feedback collected None

100%

Telemetry precision

94%+

Flight accuracy

3

LLM pipeline modules

Active

Project status

To appear at UIST '26 · Detroit, MI