AI copies do not multiply proof
The plain-language rule is brutal: count evidence roots, not AI voices. If 32 AI systems repeat the same thin source, the system can look independent while its proof stays narrow. Aggregators need to track ancestry and dependence, not just report quantity or similarity.
What the source reports
Marc Bara formalizes an epistemic Sybil problem for multi-part AI systems. Epistemic means about knowledge. A Sybil problem is when many apparent identities hide one source. The paper argues that another AI report is not another observation. Reports may come from the same evidence. Separate evidence can also lead to similar reports. The study used more than 20,000 controlled report and extraction calls with AI models. When one evidence root stayed fixed and report count rose from 1 to 32, naive posterior coverage fell from 0.940 to 0.263. When true evidence roots rose from 1 to 16, the gap closed. At k = 16, aggregators were statistically indistinguishable. The design rule is to track evidence ancestry and dependence.
Original source
- Title
- Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence
- Author
- arXiv
- Publication
- arXiv