Android Andrea fails to boost museum visitor acceptance with ChatGPT emotion simulation
73 visitors couldn't tell whether the robot was faking emotions via ChatGPT 4.1 or not.
In a follow-up experiment at a German museum, researchers placed the android robot Andrea autonomously for six consecutive days, engaging visitors in multilingual conversations about exhibits. The robot featured a humanlike, gender-ambiguous design with a slightly artificial voice. Three experimental conditions were tested: (1) no emotion simulation, (2) emotions generated by ChatGPT 4.1, and (3) emotions modeled by the WASABI simulation architecture. A total of 73 visitors completed an extended TAM2 questionnaire to evaluate their experience.
Statistical analysis revealed that neither ChatGPT 4.1 nor WASABI-based emotion simulation produced any significant positive effect on visitor acceptance, likability, or perceived humanness. Participants showed no conscious awareness of the emotional variations, suggesting that current emotion-generation methods for humanoid robots may not translate to improved real-world interactions. The findings challenge assumptions that adding emotional AI to social robots will automatically enhance user experience, especially in public-facing deployments like museums.
- Android Andrea ran fully autonomously for 6 days at a German museum, holding multi-lingual conversations about exhibits.
- Three conditions were tested: no emotion, ChatGPT 4.1-driven emotions, and the WASABI emotion architecture.
- Survey results from 73 visitors showed no statistically significant improvement in acceptance or detectability of simulated emotions.
Why It Matters
Emotion AI in humanoid robots may not improve public acceptance, questioning assumptions in social robotics design.