Contemplative Superalignment
Ruben Laukkonen, Fionn Inglis, Shamil Chandaria, Lars Sandved-Smith, Edmundo Lopez-Sola, Jakob Hohwy, Jonathan Gold, Adam Elwood
Artificial General Intelligence 2026 DOI: 10.1007/978-3-032-00686-8_31 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Empirical study Peer reviewed |
|---|---|
| Population | Artificial intelligence systems |
| Topics | Buddhism Meditation |
| Keywords | Artificial intelligence Alignment Large language models Neural networks Machine learning Compassion Non-duality Contemplative science Neurophenomenology |
| Citations | 1 |
| Key points | Prompting AI with contemplative principles significantly improves alignment benchmark performance and cooperation in a game-theoretic task. |
Abstract
As artificial intelligence (AI) improves, current alignment strategies may falter in the face of unpredictable self-improvement and the sheer complexity of AI. Rather than trying to control behavior, we show how four principles from contemplative traditions can help intrinsically align (super) intelligence. First, mindfulness enables self-monitoring and recalibration of emergent subgoals. Second, emptiness forestalls dogmatic goal fixation and relaxes rigid priors. Third, non-duality dissolves adversarial self–other boundaries. Fourth, boundless care motivates the universal reduction of suffering. We find that prompting AI to reflect on these principles improves performance on the AILuminate Benchmark ( d = .96) and boosts cooperation and joint-reward on the Iterated Prisoner’s Dilemma task ( d = 7 +). We also show how active inference offers parameters for integrating contemplative wisdom deeper into the architecture and world models of AI. This interdisciplinary approach offers a resilient alternative to brittle control schemes and may be the first empirical test of ‘ancient wisdom’.