Ambivalent Personas: Simulating Realistic Users with AI
Why polished AI personas fail and how targeted prompts can simulate realistic contradictions in user behavior, including a practical example.
Meet “Perfect Paul”: 28, obsessed with efficiency, drinks only fair-trade coffee and uses your app exactly as intended every day. The problem is that Paul does not exist.
Real people contradict themselves. We say privacy matters, then upload half our lives to Instagram. We want to eat better, then order pizza after a stressful day. Those contradictions make us human — and make design harder.
AI personas often come out too smooth. Language models are optimized for coherence, while people are not. Realistic personas therefore need explicit contradictions.
Why polished personas are dangerous
If you design for Paul, you design past real life. The classic example is the privacy paradox: people say privacy matters, then spend hours on social platforms. If your design listens only to what people say, it may be ignored. If it watches only what they do, it can feel intrusive. The useful insight sits in the reason for that gap.
Teaching an AI to contradict itself
You can tell the model to stop producing a typical persona and ask for conflicts instead.
1. Gather real data
Start with numbers from your own research, for example:
- “70% of our users say privacy matters to them.”
- “65% use Facebook every day.”
2. Use an ambivalence prompt
Tell the model to build that conflict into every persona. This is the prompt I use:
Create exactly three realistic personas for this segment:
• Mobile-only users
• Over 45 years old
• Live in a rural area
• Say: “Privacy matters to me”
• Do: Use social media every day
For EACH persona:
1. Include at least two hard contradictions between what they say and what they do.
2. Write one quote that reveals the internal conflict, for example: “I know it is not ideal, but ...”
3. Keep it short, no more than 150 words.
Return the result as a Markdown table.What the prompt produces
Instead of Perfect Paul, you might get:
Lisa, 50, teacher.
- Conflict: She worries about data theft but posts photos of her grandchild every Sunday because otherwise she loses touch with the family.
- Quote: “I know Facebook collects too much data. But without it I never hear when the village festival is happening.”
Lisa's social pressure outweighs her privacy concern. That gives a design team something concrete to discuss: how could the product create a sense of safety without cutting her off?
These personas are still synthetic. They come from a model, not from real conversations. They are useful in workshops and as discussion starters, but they do not replace an interview with a real Lisa.
Stop looking for perfect users
Perfect users do not exist. Look for the cracks, trade-offs and small contradictions people allow themselves. If an AI persona irritates you slightly because it behaves inconsistently, it may finally be useful.