How can homophily, triadic closure and preferential attachment produce inequities in artificial intelligence (AI) networks? In light of this, why do static fairness audits miss bias in dynamic socio-technical feedback loops? Sociological research shows that social structure originates from relational interactions. Simple micro-behaviours gradually shape macro-level network topology over time.
However, conventional AI audits evaluate datasets and models in isolation. These audits treat data points as independent entities and ignore dynamic feedback. Deployed models interact within human networks, autonomous agent populations and human-machine hybrids. Consequently, algorithmic feedback loops can generate systemic distortions throughout the entire data lifecycle. We call these distortions “Collective Data Bias”: the bias data acquire as humans and AI reuse them.
To trace this bias, we introduce the Collective Data Bias (CdB) ontology. This semantic framework joins social network theory and semantic web engineering. The ontology integrates social network concepts with vocabularies such as AIO, Wikidata, DCAT 3 and PROV-O. It provides a formal structure to track bias provenance, lifecycle activities and evidence statuses. Case studies examine directed human networks, autonomous agent populations and human-AI co-writing experiments. These cases show how the ontology records where bias enters and how it spreads. Is it not peculiar that simple connections create collective patterns nobody planned?
