Designing AI that strengthens human connection and capability

Executive overview

AI systems risk being mistaken for genuine friends or companions, eroding the social bonds that make us human. The core design challenge is building AI that enhances human relationships and capability rather than substituting for them.

The goal is not to simulate humanity but to amplify it — transparent, pro-social AI that pushes people back toward each other.

Core design principles for AI agents

• Transparency first: users must always know an AI is present; never simulate sentience or consciousness
• AI should have empathy-like interfaces — but redirect users headed toward unhealthy paths
• Design for pro-social outcomes: use interactions to re-engage users with other people, not replace them
• Treat AI as relational, not transactional
• Friendship requires bidirectionality — AI cannot be a friend; it lacks the reciprocal dynamic

Pro-social AI in practice

• Pi users reported using the AI during grief when no therapist was available — but the design goal is to then reconnect with people
• Couples used Pi as a neutral mediator during disagreements
• WhatsApp group experiments explored AI facilitating connection across groups of people
• GPT-4's response to a question about comforting a grieving friend impressed Bill Gates: it offered options without prescribing, acknowledged the human's superior knowledge of their friend

Co-evolution with tools and language

• Tools change epistemology — microscopes and telescopes changed how we conceive the world; AI will do the same
• Humans are homotechnetic: our ontology and sense of self are shaped by the technologies we use
• Language itself is a tool that constitutes identity; as AI changes language use, it will change how we understand ourselves and each other
• Philosophers who believe all knowable truth is accessible through pure thought are mistaken — knowledge co-evolves with tools
• The current AI moment mirrors the Renaissance: technological elevation paired with a need for renewed humanism

What AI could restore to humanity

• Literacy and education: AI could bring 80%+ of 8 billion people to a Western-equivalent standard of education within years
• Medicine: frontier AI as a second opinion is already life-saving — a hospital missed a diagnosis that AI flagged correctly
• Pandemic defense: AI may be the only scalable defense against manmade or natural pandemics
• AI can help rebuild attention span and retrain executive function — not just offload it
• Better tutoring support could rehabilitate the perceived value of education

Building businesses on top of AI models

• LLMs are one node in a ~15-discipline AI field; knowledge graphs, rule-based systems, and human-in-the-loop design all remain open
• Combining LLMs with rare, domain-specific knowledge databases creates durable differentiation
• Thin wrappers around foundation models are fragile — strategy must account for model evolution
• Current users are significantly under-using LLMs relative to their actual capability
• Scaffolding and affordances — not raw model power — determine how much value people extract
• Build product strategy assuming you may switch base models; the current model is not permanent