The reference set decides what the model can see.

A reference makes some differences visible

A model cannot recover variation that its reference made invisible. That is the simplest lesson I carry from genomics into AI governance.

A reference genome is a coordinate system used to align, compare and describe biological sequences. It is extraordinarily useful, but it is not a universal human genome. When a workflow treats one reference as the definition of normal, differences outside that representation may become harder to map, classify or even notice. The same pattern appears in AI: the baseline dataset, taxonomy or benchmark quietly determines what counts as expected, anomalous or absent.

Genomics learned to question the coordinate system

Two peer-reviewed advances make the teaching point concrete. In 2022, the Telomere-to-Telomere Consortium published T2T-CHM13, addressing long-standing gaps in the human reference and providing a complete sequence for the autosomes and chromosome X in that assembly. In 2023, the Human Pangenome Reference Consortium published a draft reference built from 47 phased, diploid assemblies from genetically diverse individuals. Instead of forcing every comparison onto one path, a pangenome can represent many paths through human variation.

A separate 2022 Nature Genetics study that I coauthored analyzed plasma proteins in participants of European and African ancestry. The published findings showed the value of studying diverse ancestry groups for understanding the genetic regulation of molecular traits. That peer-reviewed result does not prove that every diverse dataset is sufficient or that every downstream model is fair. It does reinforce a durable methodological lesson: representation and reference design belong inside the scientific question, not outside it as invisible plumbing.

Reference choice is a governance decision

Before evaluating an AI system, document the reference set it uses. Who is represented? Which versions, time periods and environments are included? What was excluded? Which groups or conditions have sparse coverage? What happens when an input falls outside the known coordinate system?

These are governance questions because reference choices shape who carries uncertainty. A system may look accurate in aggregate while concentrating mapping errors, low confidence or manual review on people who are poorly represented by the baseline. An audit should therefore inspect not only the final prediction, but also coverage, missingness, calibration and escalation across relevant groups and operating conditions.

Test the reference, not only the model

I would run a controlled comparison that holds the task, model, tools and review budget constant while changing the reference design: a single established reference; a broader multi-reference or graph-based representation; and a deliberately held-out population or operating condition.

Measure error distribution, coverage, calibration, unresolved cases and reviewer burden. Record whether a broader reference changes who benefits, who is sent to manual review and where uncertainty remains. Include negative controls so that added complexity is not mistaken for improvement.

This is a proposed evaluation, not a report of new results. Its purpose is to expose a hidden dependency. The model is only one part of the system. If we want accountable AI, we must be able to explain the coordinate system against which its world becomes visible.