Toward an Engineering of Science: Rebalancing Generation and Verification in the Age of AI
A new academic paper submitted to arXiv addresses the structural risks posed by AI-generated scientific artifacts, termed 'epistemic pollution.' The author argues that while AI significantly lowers the cost of generating plausible but potentially unreliable research papers, reviews, and surveys, it does not equally reduce the cost of verifying them. This imbalance threatens scientific integrity, as the traditional infrastructure relied on high generation costs as a natural filter. To resolve this, the paper proposes treating scientific infrastructure as an engineering problem. It introduces 'blueprints,' a novel framework where research claims, evidence, and assumptions are structured as typed graph components rather than compressed prose. This approach shifts effort toward upfront generation to enable cheaper, localized, and distributed verification downstream. A proof-of-concept prototype demonstrates the feasibility of this method. The work suggests a fundamental redesign of how scientific knowledge is produced and validated in the AI era, aiming to restore balance between creation and scrutiny to prevent the accumulation of unreliable information.
Wire timeline
Toward an Engineering of Science: Rebalancing Generation and Verification in the Age of AI
A new academic paper submitted to arXiv addresses the structural risks posed by AI-generated scientific artifacts, termed 'epistemic pollution.' The author argues that while AI significantly lowers the cost of generating plausible but potentially unreliable research papers, reviews, and surveys, it does not equally reduce the cost of verifying them. This imbalance threatens scientific integrity, as the traditional infrastructure relied on high generation costs as a natural filter. To resolve this, the paper proposes treating scientific infrastructure as an engineering problem. It introduces 'blueprints,' a novel framework where research claims, evidence, and assumptions are structured as typed graph components rather than compressed prose. This approach shifts effort toward upfront generation to enable cheaper, localized, and distributed verification downstream. A proof-of-concept prototype demonstrates the feasibility of this method. The work suggests a fundamental redesign of how scientific knowledge is produced and validated in the AI era, aiming to restore balance between creation and scrutiny to prevent the accumulation of unreliable information.
cs.AI updates on arXiv.org