About This Project
An interactive exploration of AI alignment philosophy
Project Overview
The AI Alignment Comparator is an interactive research tool that demonstrates how different AI alignment philosophies shape AI behaviour in practice. By comparing responses from three distinct frameworks to identical mental health scenarios, users can directly observe how alignment choices manifest in real interactions.
This project was developed for RELI E-1730: Mindfulness, AI, and Ethicsat Harvard Extension School (Spring 2026), examining the intersection of AI alignment approaches and Buddhist ethics principles.
The Three Frameworks
Distinct philosophies for making AI safe and beneficial
Humanist Superintelligence (HSI)
Developed by Microsoft and articulated by Mustafa Suleyman, HSI emphasises containment as the primary safety mechanism. AI operates within strictly defined domains, with mandatory human oversight and transparent, interpretable decision-making. The AI positions itself as a subordinate tool rather than an autonomous agent.
Constitutional AI
Developed by Anthropic and informed by Amanda Askell's research, Constitutional AI emphasises character as the foundation of safety. Through training, the AI internalises values that guide behaviour. The "brilliant friend" model prioritises genuine engagement, honest feedback over sycophancy, and psychological stability under pressure.
Integrated Approach (Hybrid)
A synthesised framework that combines containment architecture with character training. Clear boundaries define the scope of operation (from HSI), while genuine character determines how the AI engages within those bounds (from Constitutional). This creates "bounded authenticity" - genuine presence within defined limits.
Buddhist Ethics Integration
Ancient wisdom applied to contemporary AI challenges
The reflection questions integrated into the comparison experience draw on Buddhist ethical frameworks, particularly Peter D. Hershock's work on technology and Buddhist principles:
- Attention quality: How does each framework shape what we notice and how we engage with the interaction?
- Relational dynamics: What kind of relationship does each framework foster between human and AI?
- Compassion expression: How does genuine care manifest differently across alignment approaches?
- Appropriate boundaries: When are limits supportive, and when do they become barriers to genuine connection?
Technical Implementation
Modern web technologies for research reproducibility
Frontend
- Next.js 15 (App Router)
- React 19
- TypeScript (strict mode)
- Tailwind CSS + shadcn/ui
- Framer Motion
- Recharts
Backend
- Vercel AI SDK 6.x
- Claude Sonnet 4 (Anthropic)
- Edge runtime (streaming)
- Multiplexed NDJSON
- Turborepo monorepo
The architecture supports parallel streaming responses from all three frameworks simultaneously, providing real-time comparison as responses generate. System prompts are versioned for research reproducibility, and reasoning traces are captured for analysis.
About the Author
Cian O'Sullivan is a dual MSc student in Psychology and Neuroscience of Mental Health (King's College London) and Psychology (Harvard Extension School). His research focuses on limerence using multilevel network analysis, exploring how intense romantic preoccupation emerges from the interaction of psychological, social, and biological factors.
This project bridges academic research in AI ethics with practical software development, demonstrating how alignment philosophy translates into observable differences in AI behaviour.