Guest lecture: AI personality and user perception

Optional aside

Not examinable. These are guest lecture slides on Canvas and in Guest_Lecture_2.pdf. The week 12 slides name a guest speaker, Daria Podrezova, but the extracted slides contain no content from that talk.

Intuition

People treat talking machines like people. So the way an assistant sounds and apologises changes how much we trust it when it fails.

The research gap: personality design needs an integrated view

The talk relies on the Computers Are Social Actors (CASA) paradigm (Nass et al., 1994). AI personality traits shape user perceptions and emotional responses, but existing research is fragmented. Conversational AI personality design needs an integrated understanding of perceptions, contextual expectations and behavioural responses over time. The speaker frames it as moving from HCI (how people interact with technology) to usability engineering (designing and evaluating systems).

Project 1: voice assistant personality and failure

Related work: shaky voices increase empathy, whispering intimacy, gender affects persuasiveness, trust and rapport, perceived age affects familiarity and relaxation (with the risk of ageism). Response styles: humour (affiliative, aggressive, self-enhancing, self-defeating) and apologies (serious or casual, taking or shifting blame). Most commercial voice assistants are gendered female by default.

Design: two voices, Nexa and Nova (perceived older), and three failure responses (apologetic, humorous, no acknowledgement), six task categories, 60 participants (30F, 30M, aged 18 to 52), with measures of empathy, trust (ability, benevolence, integrity) and service recovery satisfaction. Results: Nova evoked maternal and kinship feelings, especially in male participants. Humour was witty to some and ridiculing to others. Apology read as reliable and polite, but also robotic. No acknowledgement led to confusion and frustration. Conclusions: sincere acknowledgement helps recovery, depends on gender and context, and a context-based balance of apology and humour may work best.

Projects 2 and 3

Project 2 studies flexible user-driven configuration: a platform that lets users adjust a conversational agent’s personality during a conversation with sliders across eight dimensions (decency, profoundness, instability, vibrancy, engagement, neuroticism, serviceability, subservience; Kovacevi et al.). Converged classes of use were problem-focused guidance, affective coping and reflective evaluation, with serviceability dominant. Project 3 is a 21-day field study of an AI voice health coach to motivate sedentary users. Contributions of the programme: a unified view of AI personality across perception, expectation and behaviour, design guidelines, and sustained real-world impact.

Recall

What did Project 1 conclude about apology and humour?

Sincere acknowledgement leads to better failure recovery, and a context-based balance of apology and humour may work best.