Halo · UCRE / XRAY
Users of AI design tools rarely notice the bias in what those tools produce. Halo is a companion that learns which biases you care about, asks before you generate, and flags what it finds after.
People using AI design tools are trying to make something fast. Canva AI gives you three outputs from a sentence; you compare them, pick the best, and move on. The question we started with was whether anyone in that loop stops to check the image for bias.
For a semester project in User-Centered Research and Evaluation, a team of four ran 8 think-aloud sessions over three days. Each participant completed three design prompts in Canva AI while narrating what they were doing and thinking. We synthesized what we heard through empathy mapping, affinity clustering, and a user-flow model that set the current experience alongside where Halo would sit inside it.
Most participants noticed bias that was obvious: a "professional" setting with no women, a "diverse" group that looked narrowly defined. They would re-prompt once and usually get something better. What they did not do was look for bias before accepting an output. When something was subtle, they didn't name it. When fixing it would take more than one extra prompt, they let it go.
Outside the tool, they had the knowledge. They had heard that AI could produce biased images, that careless prompting could make it worse, that they had some responsibility. One participant said they felt "a strong responsibility to correct the AI outputs." But knowing something in general doesn't mean you notice it in the specific image you're about to use. The gap wasn't awareness. It was the moment between generating and accepting, which the current tool made essentially invisible.
"Honestly, I wouldn't have even noticed it."
They knew AI was biased. They didn't see it in the image they were about to use.
If users won't go looking for bias on their own, putting a tool somewhere they'd have to seek it out was the wrong move. The question was where in the experience Halo could actually be.
Halo doesn't add a new place to go. It's already where the work happens. The prompt bar is where every session starts, and that's where the bias often starts too.
Even inside the flow, there were two moments to work with: before you generate (vague prompts are where most bias gets introduced) and after (outputs that slipped through anyway). We built both.
Setup lets users rank which biases matter most to them: representation and diversity, cultural sensitivity, gender stereotypes, body and appearance, socioeconomic cues. The Nudge asks three or four clarifying questions tied to those priorities, then rewrites the prompt with the constraints included. The Flag scans what was generated and breaks it down by area: gender split, age distribution, disability representation, cultural context, racial and ethnic diversity, socioeconomic cues. Every result is framed as coming from training data, not from the user's intent. The Nudge handles bias the user knows to watch for. The Flag handles what gets through anyway, and answers the question that kept surfacing in the research: I know AI can be biased, but is this image biased?
I illustrated Halo. the "halo effect" in psychology is when one good quality makes you overlook the bad ones. the name is a small joke about exactly what it's trying to fix.
research poster presented to CMU HCI faculty and classmates
prototype: Setup, Nudge, and Flag stages overlaid on Canva AI
and critiqued by faculty and peers; friction in the Nudge flow sharpened the case for a one-tap quick mode
The critiques that came back are real ones. The Nudge adds questions before every generation, and the participants who said they'd "just use it" are exactly the ones most likely to skip it. The refined prompt can veer into checklist diversity, close to the tokenism it's meant to avoid. And inferring gender, race, age, and disability from a generated face is itself a judgment call. A one-tap "quick" mode and a study on where people actually stop are the obvious next steps.
If a user says "I'd just use it," you can tell them not to or you can be there when they do.