The Energy of Answers

A calm, data-led reflection on what AI costs to run.

An interactive web experience that guides users through a thoughtful sequence of prompts and visuals—revealing how “frictionless” AI use maps to real infrastructure, real energy demand, and subtle social trade-offs.

This project is intentionally reflective. It is designed to provoke better questions—not to prove causal claims.

---

Target Audience

• Everyday AI users who want more grounded sense of AI’s hidden footprint.
• Students, educators, and workshop facilitators exploring tech + society through data storytelling.
• Designers, builders, and decision-makers who want a simple experience to spark better conversations about sustainability and wellbeing.

---

Problem Statement

AI is becoming a daily utility. Yet its costs are often:

• Invisible (compute, data centers, electricity)
• Abstract (charts without lived context)
• Easy to misread (correlation mistaken for causation)
• Socially consequential (tools that reduce friction can also reshape habits, connection, and support)

This project addresses that gap by pairing calm storytelling with public statistics, helping users feel the tension between:

• Convenience ↔ infrastructure
• Scale ↔ sustainability
• Connection ↔ community

---

Key Features (Summary)

• Guided storytelling flow: question → answer → visual → insight
• Progressive disclosure: simple framing first, then layered “reveal” visuals
• Reflection inputs: users respond with feeling/expectation choices
• End-of-flow personalized reflection: a short tailored reflection generated from the user’s responses
• Curated indicator selection: indicators were manually screened and selected from a large public dataset catalog for narrative fit

---

Wellbeing Dimensions Covered (Current Scope)

This version focuses primarily on:

1) Environmental wellbeing (primary)

Explores the relationship between digital life, electricity demand, and the infrastructure needed to support modern AI.

2) Social wellbeing (supporting lens)

Uses the social dimension to frame a key question: as tools become more helpful and ever-present, what happens to community, attention, and human support?

With more time, the project expands into the remaining wellbeing dimensions (physical, mental, economic, cultural) using the same rigorous “question → answer → visual → insight” approach.

---

AI Feature Implemented (Current)

AI is used at the end of the flow to generate a tailored reflection.

• User inputs collected (categorical): the app captures the user’s responses across a small set of reflection questions.
• One OpenAI call: those responses are inserted into a structured system prompt to generate a reflection that is shown to the user to encourage thought-provoking conversations and drive change.

---

Data Sources (High Level)

Core Statistics

Most baseline indicators and time series visuals are built from Gapminder’s Systema Globalis dataset (DDFcsv), a comparable collection of public statistics used across the Gapminder ecosystem.

Repo:

• https://github.com/open-numbers/ddf--gapminder--systema_globalis

Example files used:

• ddf--datapoints--electricity_use_per_person--by--geo--time.csv
• ddf--datapoints--electricity_generation_total--by--geo--time.csv
• ddf--datapoints--urban_population--by--geo--time.csv
• ddf--datapoints--internet_users--by--geo--time.csv
• ddf--datapoints--cell_phones_per_100_people--by--geo--time.csv

AI / Technology Context Datasets

Used to contextualize the recent acceleration of AI adoption and investment:

• OWID: Share of AI job postings
  https://ourworldindata.org/grapher/share-artificial-intelligence-job-postings
• OWID: Private investment in AI
  https://ourworldindata.org/grapher/private-investment-in-artificial-intelligence
• IMF: AI Preparedness Index (AIPI); To be incorporated in a later version. https://www.imf.org/external/datamapper/AI_PI@AIPI/ADVEC/EME/LIC

Anchoring References (Connectivity Context)

Used to ground long-run shifts in global connectivity:

• ITU Measuring Digital Development (2019)
  https://www.itu.int/hub/publication/d-ind-ict_mdd-2019/
• ITU Measuring Digital Development (2022)
  https://www.itu.int/hub/publication/d-ind-ict_mdd-2022/

---

Methodology

• Manual indicator screening: we reviewed the Systema Globalis concepts catalog (ddf--concepts.csv) and manually shortlisted indicators that could form meaningful pairs by wellbeing dimension (positive/negative tensions).
• Minimal preprocessing: datasets were used largely as provided.
• Aggregation: where needed (e.g., AI job postings), we aggregated to form a world-level proxy.
• Forecasting for missing recent years: when core indicators lagged, ARIMA was used to estimate the most recent year(s) so the narrative could include recent windows where relevant.
• Dual-axis charts by design: some indicators differ in units and scale (e.g., per-person use vs total generation). Dual axes preserve interpretability without forcing normalization.

---

Limitations & Considerations

• Correlation ≠ causation: the experience does not run causal inference or statistically powered hypothesis testing.
• Confounders exist: trends may reflect electrification, GDP, population growth, policy, grid mix, measurement changes, or reporting delays.
• Averages can hide real lives: world/country averages can mask inequality and local variation.
• AI coverage is uneven: some AI datasets have partial coverage or are more representative of some regions than others.
• Forecasted years are estimates: ARIMA outputs are model-based extensions, not official measurements.

---

Technology Stack & Key Libraries

Frontend

• Next.js 15.1.4 - React framework with App Router for server-side rendering and routing
• React 19.0.0 - UI library for building interactive components
• TypeScript 5.9.3 - Type safety and enhanced developer experience

Visualization

• Recharts 3.7.0 - Chart library for creating responsive data visualizations (AreaChart, LineChart, etc.)

Data Processing

• Python 3.13 (with virtual environment)
• pandas (>=2.0.0) - Data manipulation and analysis for processing Gapminder datasets
• numpy - Numerical computing (used in Jupyter notebooks for data analysis)
• statsmodels - ARIMA forecasting for estimating missing recent years in time series data
• requests (>=2.28.0) - HTTP library for fetching data from external sources

Backend/API

• Next.js API Routes - Serverless API endpoints for handling reflection generation requests

AI

• OpenAI API (via openai package v6.16.0) - GPT-4o model for generating personalized end-of-flow reflections

Styling

• CSS Modules - Component-scoped styling for maintainable UI components
• CSS Custom Properties - Theme variables supporting light/dark mode preferences

Development Tools

• ESLint - Code linting with Next.js configuration
• TypeScript - Static type checking and improved IDE support

---

Setup Instructions

Prerequisites

• Node.js 18.x or later
• Python 3.13 (or 3.9+)
• npm or yarn package manager
• OpenAI API key (for AI reflection generation)

Installation Steps

1. Clone the repository (if applicable) or navigate to the project directory.

2. Install Node.js dependencies:

   npm install

   This will install all frontend dependencies including Next.js, React, TypeScript, Recharts, and OpenAI SDK.

3. Set up Python environment (optional, for data processing scripts):

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt

4. Configure environment variables: Create a .env.local file in the root directory:

   OPENAI_API_KEY=your_openai_api_key_here

5. Run the development server:

   npm run dev

   The application will be available at http://localhost:3000.

Notes

• The Python environment is only needed if you plan to run data processing scripts or Jupyter notebooks.
• The OpenAI API key is required for the reflection generation feature to work. Without it, the app will run but the AI reflection endpoint will return an error.
• Data files are already included in the data/ directory, so no additional data download is required to run the web application.

---

User Guide

1. Begin the experience (Environmental dimension).
2. Read short prompts and respond with a choice when signaled.
3. Progress through a sequence of visuals and insights that build complexity gradually.
4. Near the end, the app generates a tailored reflection based on your responses.
5. At the end, you can click View Sources & Methodology for transparency and references.
6. Continue to the next section/dimension (as available).

---

What Actions Users Can Take (Based on Insights)

This experience is designed to end with agency, not anxiety. Examples:

• Use AI more intentionally (avoid reflex prompts when other tools fit).
• Prefer lighter interactions when possible (fewer retries, fewer image generations).
• Use traditional search for simple factual lookups.
• Support transparency and sustainability norms (energy labels, reporting, cleaner grids).
• Reinvest in community: replace some “AI-as-companion” moments with human connection.

---

What We Would Build Next (With More Time)

• Complete the remaining wellbeing dimensions (physical, mental, economic, cultural) with the same bespoke “question → answer → visual → insight” craftsmanship.
• Filtering & comparison modes: allow users to view a single country, compare countries, or switch between world aggregation and user-selected groupings.
• Model-choice education through experience: add a “double reflection” ending:
  • Generate two reflections from the same user inputs—one using a large hosted model (LLM) and one using a smaller local model (SLM) (e.g., via Ollama / llama.cpp).
  • Present both reflections without labeling which is which.
  • Then reveal the difference in estimated cost/energy footprint and discuss when SLMs are “good enough,” encouraging experimentation with lower-impact defaults.
• Mobile adaptation: adapt the web experience for a mobile-first flow.

---

Privacy

User responses are collected to generate the end-of-flow reflection (and may be sent as part of an OpenAI API call). We do not use responses for profiling, and we do not publish or sell response data.

---

License

• Data: Gapminder Systema Globalis is provided under CC BY 4.0 (see dataset repo for details).