--- title: 'A Paradigm Shift in AI Safety' subtitle: 'Make protecting life part of how AI improves.' slug: 'paradigm-shift-in-ai-safety' date: 2026-10-09 type: 'essay' status: 'published' tags: ['enviroai', 'recursive-self-improvement', 'incorruptible-grader', 'environmental-superintelligence', 'accessible'] abstract: 'A one-page strategy brief introducing RESI (recursive environmental self-improvement) as a proposed, testable contribution to AI safety: let AI improve its methods inside a loop of observing, acting within limits and independent verification against physical outcomes, while human-set purpose and limits stay outside the system''s authority to change. The living world serves as the external check—does greater capability actually deliver greater protection?' license: 'CC-BY-4.0' author: 'Jed Anderson' co_authors: [] canonical_url: 'https://jedanderson.org/essays/paradigm-shift-in-ai-safety' pdf: '/pdfs/paradigm-shift-in-ai-safety.pdf' hero_image: '/images/paradigm-shift-in-ai-safety-hero.jpg' hero_image_alt: 'One-page strategy brief. A dark navy banner across the top carries a small ''A proposed strategy'' label, the title A Paradigm Shift in AI Safety in large white serif type, and the line ''Make protecting life part of how AI improves,'' with ''Introducing RESI—Recursive Environmental Self-Improvement, a path toward Environmental Superintelligence'' beneath. On a cream background below, a diagram labelled ''The safety mechanism'' shows a black bar of human-set purpose and limits above a loop of four boxes—Observe, Act within limits, Independently verify (highlighted in blue), Improve methods—beside a pull quote; a dark ''Guardrail'' band, three short columns on why to begin in the living world, the external test question, and a footer pointing to the essay The Planet Is the Benchmark with a QR code.' supporting_files: [] related_essay: '/essays/planet-is-the-benchmark' --- *A proposed strategy. Introducing RESI—Recursive Environmental Self-Improvement, a path toward Environmental Superintelligence.* ## As AI becomes more capable, how do we keep its improvement accountable to human purposes and real-world consequences? **RESI proposes a starting point: make protecting life part of how AI improves.** Begin in the living world, where independently measured outcomes can test whether greater capability delivers greater protection. Humans set the purpose and limits. The system improves its methods—not its mandate. ## The safety mechanism **Human-set purpose and limits.** Set by people. Outside the system’s authority to change. Inside those limits, the system improves its methods in a loop: 1. **Observe** 2. **Act within limits** 3. **Independently verify**—not self-graded 4. **Improve methods** Independently verified results inform the next iteration. The loop learns. The learner improves the loop. Real-world evidence informs the next improvement. > *“Don’t wait for greater intelligence and then ask it to protect life. Make protecting life part of how it improves.”* **AI already helps develop AI. RESI asks what should guide that improvement.** Direct some of it outward, toward the living world, and build an intelligence that learns: - not only to answer, but to observe; - not only to act, but to verify; - not only to improve, but to improve its capacity to protect life. ## The guardrail Under human oversight, the system may improve its methods—not redefine its purpose, expand its authority, or lower its standard of proof. ## Why begin in the living world - **An external check.** Physical outcomes provide an external check on the system’s claims. - **Measured progress.** Progress must be demonstrated through independently measured outcomes. - **A protective purpose.** The purpose of greater capability is protective rather than extractive. ## The external test > Does it understand the Earth better—and does that understanding allow life to flourish more securely? Read the essay: [The Planet Is the Benchmark](/essays/planet-is-the-benchmark). *A proposed, testable contribution to AI safety—not a claim that environmental performance alone proves an AI system safe.*