People increasingly turn to ChatGPT for health questions and self-diagnosis. The problem: large language models can generate incorrect information with high linguistic confidence, a phenomenon known as hallucinations. When users trust these confident-sounding answers uncritically, known as overreliance, it can negatively affect their health decisions.
Previous research pointed to color as a promising way to signal uncertainty in AI interfaces, but different approaches had only been studied separately. No study had compared multiple color variants directly within ChatGPT's interface in a health context. That was the gap this study set out to explore.
1. How do different ways of implementing color in ChatGPT's interface affect users' critical evaluation of health-related AI answers during self-diagnosis?
2. Which color variant do users prefer for identifying hallucinations and reducing overreliance?
We used a research-through-design method, an established approach in HCI where prototypes act as both the research tool and the object of study. This let us explore not just whether color helped, but how users actually responded to different implementations.
The study combined two methods:
We built on existing research (Leiser et al.'s HILL framework), which identified color-marking as one of the most significant design elements for surfacing potential hallucinations, and focused our study specifically on that element.
We began by grounding the study in existing research on how large language models generate hallucinations, and why users tend to over-trust confident AI answers. We built on Leiser et al.'s HILL framework, which identified color-marking as one of the most significant design elements for surfacing potential hallucinations.
Alongside the literature, we ran qualitative in-depth interviews with our target group (young adults, 18–29, without medical training) to understand when and why they turn to ChatGPT for health questions. A key early insight emerged: overreliance is situational rather than a fixed habit, and simply being aware that hallucinations exist did not automatically lead people to check information critically.
The interviews helped define the core problem: users needed a way to see which parts of an AI answer were uncertain, without being overwhelmed. From the five design elements identified in prior research, we focused specifically on color as a way to signal uncertainty.
We developed three prototypes modeled on ChatGPT's interface, each applying color differently to mark uncertain information in an AI health response. Each variant was designed to test which approach best supported critical evaluation.
We tested the three prototypes with six participants (18-29, without medical training) using think-aloud and observation. Testing explored how each color variant affected whether participants noticed uncertain information, how they valued it, and which variant they found clearest.
All three variants increased participants' awareness of uncertain information and reduced overreliance, though awareness alone didn't always lead them to double-check. The majority preferred Interface 3, uncertain information marked in red, for its clarity and readability.
The study surfaced several insights about using color to communicate AI uncertainty:
The biggest lesson came from the prototype itself. I invested heavily in the UI of a medium-fidelity prototype before testing any of it, which means if testing reveals problems, some of that polish may turn out to have been premature. Next time I'd validate the core flow at a lower fidelity first, and only invest in detailed UI once the structure holds up. Fidelity should follow confidence, which means the more certain I am that something works, the more polish it's worth.
A second takeaway is restraint is a design decision in itself. Choosing a familiar map over impressive technology felt almost too simple at first, but the research showed that accessibility and familiarity matter more to real visitors than more advanced technology.
The study points to clear next steps: a larger sample to strengthen the findings, and testing yellow against red to see which better communicates uncertainty.
Accessibility is also key and color alone shouldn't be the only signal, especially for users with color blindness (a core WCAG principle). A further direction is exploring whether combining color with source references creates a stronger effect on critical evaluation than color alone.
Interested in a collaboration or just wanna connect?
Send me an email