AI characterCase study
The Personality Lab.
A system for building, configuring, and measuring distinct AI personas, so AI can feel like someone, not something.
Back to all workIllustrative shapes. The axis names stay unpublished until the paper is out.
- Role
- Senior UX Designer, and the only person on it
- Where
- Applied Social Media Lab, Berkman Klein Center
- Method
- 5 axes, 7-point scale, 15-prompt battery
- Status
- Research paper in progress
The question
Most bots have personality only at the surface.
Change the tone and the reasoning underneath stays the same. I wanted to know what it would take for different AI configurations to behave differently, in ways you can test and reproduce. No tool existed to do it, so I built one.
How it works
- 01BuildA detailed persona template defines who the character is before a single prompt runs.
- 02CalibrateFive behavioral axes, each scored on a 7-point scale: a true midpoint and three degrees of intensity either side.
- 03Stress-testFifteen standard prompts: problem-solving, emotional support, explanation, ethical judgment, self-doubt, and decisions.
Known characters as ground truth
To check the scale against reality, I mapped familiar personalities onto the axes first.
- Real Genius ensemble
- Socrates
- Buddha
- Nietzsche
- Marcus Aurelius
- Emerson
- Beethoven
- Bach
- Chopin
- Debussy
- The Muppets
Intellectual lineage
From stylistic dials to behavioral outcomes
I started from Anthropic’s Assistant Axis research, which measures controllable personality dimensions in language models, for a shared vocabulary. Those dimensions describe stylistic adaptation. Mine target behavioral outcomes: psychological safety and engagement in live events, without the persona drifting. I added the dimensions that were missing and landed on five, each mapped to a measurable parameter in the architecture.
- Why seven pointsPeople reliably tell about seven categories apart (Miller’s Law), and it mirrors the Likert standard: a true midpoint and three degrees of intensity on each side.
- Three layersSensors, a decision layer, and a voice synthesizer. Personality lives in the decision layer, not just the voice.
- Ground truth by contrastClusters are chosen to pull against each other. If Socrates, Nietzsche, and Marcus Aurelius don’t score meaningfully differently from the Muppets, the model is wrong. The composers test whether the axes carry over from verbal to aesthetic temperament.
Tone is aesthetic. Behavior is architectural.
- Problem-solving requests
- Emotional support
- Explaining a concept
- Ethical and judgment calls
- Self-doubt and discouragement
- Process and decision questions
Where it goes next
Personas that pass the full process can power AI-mediated civic dialogue in ongoing deliberative research. The open question: does personality configuration predict conversation quality? A colleague’s 27-tone engagement model lines up with the five axes, which suggests it might.
Historiesis: a model’s configured history shapes how it reasons about something new.