---
# System prepended metadata

title: AI Personas

---

# AI Personas: Interview Approach

## Current approach
![image alt](https://i.ibb.co/ns4NpXr5/Gemini-Generated-Image-jnnsg7jnnsg7jnns-1.png)

## The Voting Test

![image](https://i.ibb.co/Vpg9qwL8/unnamed.jpg)

### Persona Generation

![3 personas](https://i.ibb.co/pvFmRzsh/Gemini-Generated-Image-vrffevrffevrffev.png)

#### 1. Meta Persona: The Big Picture

Using general data, like census or population statistics.
- Sexual
- Age

#### 2. Tabular Personas: The Profile

Adding specific details

- **Objective**: Income, education, and job.
- **Subjective**: Political views and hobits.

#### 3. Descriptive Personas: The AI Sketch

Letting the AI do the simulation to persona's "life story" based on the data above.

### Election Day

#### 2024 Election
![2024](https://arxiv.org/html/2503.16527v1/x4.png)

#### 2020 Election
![2020](https://arxiv.org/html/2503.16527v1/x5.png)

#### 2016 Election
![2016](https://arxiv.org/html/2503.16527v1/x6.png)

### Findings

#### Postive tendency
![image](https://arxiv.org/html/2503.16527v1/x10.png)
![word cloud](https://arxiv.org/html/2503.16527v1/x11.png)

#### The "Iceberg" Problem
![image](https://arxiv.org/html/2506.06958v3/Figures/teaser.png)
- **Above the water (Surface)**: AI is good at copying how we talk and what opinions we have.
- **Below the water (Deep Thinking)**: AI doesn't actually understand why humans believe what they do. It just creates "stereotypes" instead of real, diverse people.



## The Goal: Making AI "Think" Like a Human

![image](https://i.ibb.co/0pqCvLJ6/Gemini-Generated-Image-r7rw5lr7rw5lr7rw.png)

#### Fidelity: Realism
Does the AI follow a human path to a decision, or is it just a "black box" that spits out an answer?

#### Individuality: Uniqueness
- Illusion of Consensus: AI dont like mistakes, they don't be evil
- In the real world, people disagree and have unique quirks.


### Human Reasoning

#### Causal: Cause and Effect
Understanding that one thing leads to another.
> 👩‍: "Someone is grabbing an ice cream, there must be an ice cream truck nearby."
![image](https://i.ibb.co/DP89040J/Gemini-Generated-Image-9vnjq59vnjq59vnj.png)

#### Compositional: Building Blocks
Taking old experiences and using them in new situations.
> 👩‍: "The ice cream car is surrounded by many people, there must be a line of customers"
![image](https://i.ibb.co/4gTzfxCf/Gemini-Generated-Image-xhsbp9xhsbp9xhsb.png)

### Revisable: Correct Yourself
Learning from mistakes and updating beliefs.
> 👩‍: "If I saw Conan next time, there must be something bad just happened..."
![image](https://i.ibb.co/BVdLrcxQ/Gemini-Generated-Image-rw6js9rw6js9rw6j.png)

## The GenMinds Solution
![image](https://arxiv.org/html/2506.06958v3/Figures/fig6_example.png)

#### Step 1: Structured Thought Capture - Interviewing Humans
Let AI interview real users:

> 🤖: How "Transparency" leads to "Public Safety"?
👦: It can reduce crime by aiding investigations with more transparency, which increases public safety.


**Transparency → Crime rate → Public safety**
> 🤖: Does transparency have an effect on support for surveillance?
👩‍🦰: “People will have less privacy concern if they know how data is used…
    
**Privacy Concern ← Transparency → Crime rate**

#### Step 2: Causal Belief Network - Mapping Beliefs
It builds a "web" of how that specific person thinks.

#### Step 3: Inference via Intervention - Testing Changes
> 👦: "If the government ensures that the data is used for legitimate purposes and provides explanations of those purposes, how will this specific person's opinion change?"

## References
- [Skill.md](https://github.com/search?q=repo%3Amattpocock%2Fskills%20interview&type=code)
- [LLM Generated Persona is a Promise with a Catch](https://arxiv.org/html/2503.16527v1)
- [Simulating Society Requires Simulating Thought](https://arxiv.org/html/2506.06958v3)