Can AI-generated personas build genuine empathy in product teams?

The Generating Proto-Personas study tested a prompt-engineering approach for creating proto-personas during Lean Inception product discovery with 19 participants. The efficiency gains are dramatic: creation time dropped from days to under six minutes.

But the empathy findings reveal a structured gap:

1. Cognitive empathy (understanding what the persona experiences): strongly supported. Participants could identify persona needs, behavior patterns, and contextual situations.
2. Affective empathy (feeling what the persona feels): lowest agreement among participants on Likert scale. Participants understood personas intellectually but didn't experience emotional resonance.
3. Behavioral empathy (motivation to act on persona's behalf): mixed results. Some participants were motivated to adapt the product; others were not moved to action.

The empathy gradient — cognitive > behavioral > affective — has a design explanation. LLM-generated personas excel at structured attributes: education, age, occupation, behavior patterns. These support analytical understanding. But they lack the narrative specificity, lived contradiction, and emotional texture that generates "this could be me" identification. As one participant noted, the personas had "needs that are a little different from those that were raised here" — the generalization problem that erodes emotional connection.

An unexpected benefit: because proto-personas were generated by AI rather than a human team member, participants felt more comfortable critiquing them. This lowered the barrier to discussion and produced richer refinement. The non-human origin paradoxically increased team engagement.

The LLM limitation in specialized domains was visible: legal-domain personas were generic rather than capturing the specific needs of attorney advisors vs. general legal document analysts. This echoes Why do language models struggle with historical legal cases? — legal domain specificity is a consistent LLM weakness.

The implication for design practice: AI-generated personas are effective as discussion starters and cognitive alignment tools, but insufficient as empathy generators. Teams still need techniques — narrative scenarios, real user quotes, contextual immersion — to bridge the affective gap.

The Partner Modelling Questionnaire (PMQ) provides an empirical measurement framework for how users evaluate AI dialogue partners that maps onto this empathy gradient. The PMQ identifies three factors: communicative competence/dependability (49% variance), human-likeness in communication (32%), and communicative flexibility (19%). Factor 1 (competence) maps to cognitive engagement — understanding what the agent can do. Factor 2 (human-likeness) maps closer to affective engagement — perceiving warmth, empathy, authenticity. The dominance of Factor 1 over Factor 2 mirrors the proto-persona finding: users engage cognitively before (if ever) engaging affectively (How do users mentally model dialogue agent partners?).

However, the "Computer says No" argument for against empathetic AI suggests this limitation may be the ethically correct design outcome. Genuinely empathetic response requires character knowledge — knowing whether to amplify anger for someone who doesn't stand up for themselves or de-escalate for someone who leans arrogant (Can AI give truly empathetic responses without knowing someone's character?). Proto-personas that generate cognitive understanding without affective identification may be appropriately calibrated — providing enough to inform design decisions without triggering the emotion-regulation behaviors that AI cannot responsibly perform.