If you instrument with OpenTelemetry-native packages, point the OTLP endpoint at Phoenix and let it convert gen_ai.* to OpenInference at ingest. Prefer OpenInference attributes in your own spans since those pass through untouched.

Context: Official Arize Phoenix release notes: documents the OTel GenAI semconv auto-conversion. Phoenix converts gen_ai.* attributes to OpenInference at ingest time, so traces from OpenTelemetry-native instrumentations (OpenAI, Anthropic, Google GenAI contrib packages) render with message I/O, tool calls, retrieval documents, and token counts without code changes. Spans that already carry OpenInference attributes are left untouched; conversion only fills in what is missing.