Decentralized AI's Reality Gap: 90% Theater, 10% Real Power (lightcapai.medium.com)

A new synthesis of decentralized-AI research argues that the field’s real battleground is conceptual: governance, economic incentives, ethics and legal design — not just cryptography or model math. The report aggregates taxonomies and systematizations (Yang et al.’s 2019 federated-learning categories, a 2024 SoK of blockchain‑enabled AI, the Imtidad DAIaaS blueprint) and governance/ risk artifacts (NIST AI RMF’s GOVERN/MAP/MEASURE/MANAGE pillars, ATFAA threat catalog, and the ETHOS framework’s rationality/ethics/goal‑alignment pillars and four-tier risk scheme). Empirical reviews show only ~10% of participatory projects give stakeholders real influence, and many “DeAI” projects use blockchain mainly for coordination while computation and control remain off‑chain — producing an “appearance” of decentralization rather than substantive redistribution of power. For practitioners and policymakers this matters: tokenized governance often re‑centralizes influence (token‑weighted voting) unless mitigated by mechanisms like quadratic voting, and legal/regulatory regimes (EU AI Act, GDPR) create liability and data‑immutability tensions that push architects toward hybrid designs (off‑chain storage + on‑chain proofs, ZK attestations, SSI with key‑management tradeoffs). The clear implication is that technical building blocks alone won’t deliver democratization; robust socio‑technical governance, transparent on‑/off‑chain architectures, insurance/legal entities (as ETHOS suggests), and interoperable, modular policy frameworks are required to turn decentralized AI from theater into operationally meaningful systems.