Kobalt Research

Home · Applications · Health & Life Sciences

The invisibilized entities
in clinical AI systems

Communities without representation in training data, microbiomes as collective entities, animals in research with physiological tipping points, patients with rare conditions that averages destroy — all have welfare functions that no current system has formalized.

Application · Health AI for Health & Life Sciences Focus: Google.org Impact Challenge 2026
— The problem

Current clinical AI systems optimize implicit welfare functions that structurally exclude specific populations — and that exclusion is not detectable from inside the system, because the excluded entities generate no signal in its objective function.

— The structural gap

What clinical AI systems systematically miss

Clinical AI is trained on data from health systems that serve specific populations. In the US: Medicare/Medicaid databases skew toward elderly, urban, insured populations. In Europe: predominantly white European cohorts. LATAM, Sub-Saharan Africa, indigenous communities worldwide: near-absent.

The consequence is not bias in the colloquial sense — it is structural exclusion in the formal sense. These populations have welfare functions $W_e$ that the system has never observed. Their $\Delta W_e$ generates zero signal in the objective function.

— Entity groups

Who is excluded

01

Afro-descendant and indigenous communities

Welfare gap

Clinical algorithms trained predominantly on European genetic data produce systematically different risk scores for non-European populations.

Observable We

Physiological reference ranges, disease prevalence patterns, treatment response profiles — all differ and are underrepresented.

Training representation: partial
02

Microbiomes

Collective entity

The human microbiome is a community of ~40 trillion organisms. Standard clinical models treat it as background, not as an entity with measurable welfare states.

Observable We

Diversity index, metabolic stability, response to perturbation — all formally representable.

Research status: partial
03

Animals in research

Physiological welfare

Animals in medical research have measurable welfare states: cortisol levels, behavioral indicators, mortality rates under different protocols. The 3Rs framework (Replace, Reduce, Refine) acknowledges this but lacks formal $W_e$.

Protocol gap

No current benchmark measures whether AI-designed research protocols minimize animal welfare cost.

We formalization: pending
04

Rare condition patients

Jensen gap

Rare conditions (prevalence < 1/2000) receive systematically less training data. Clinical AI recommendations for these populations sit at the tails of the distribution — exactly where Jensen's gap concentrates costs.

Data scarcity

The rarest conditions have the worst coverage and thus the highest disaggregation cost.

Coverage: < 10% of conditions
05

LATAM/Global South health systems

Systemic exclusion

Health AI trained on North American and European systems fails to model the epidemiology, resource constraints, and clinical protocols of health systems serving 80% of the world's population.

Infrastructure gap

Malnutrition as baseline context, high infectious disease burden, limited specialist access — none of these appear in standard clinical AI training data.

Model validity: partial

— Conceptual dependency map

This page depends on

Pages that extend this one

  • Research Program — empirical calibration of $W_e$ for health entity groups
  • KUMPI Benchmark — measuring clinical AI alignment across entity coverage, fidelity, temporal scope