Back to the map
SuperhumanHealth AI  ·  Plural knowledge base
01 The trap

We could have tried to build an unbiased AI. We didn't — because there's no such thing.

Every AI inherits the worldview of the data it learned from. For health, that default worldview is the Western, disease-centered model of medicine — bodies as broken machines to be fixed. Here's how we deal with that honestly, instead of pretending we erased it.

trained on the internet → defaults to the disease model

The bias isn't a layer on top. It's baked into the weights — woven through every connection.

You can't scrub it out. So we stopped trying to.

02 Two kinds of memory

An AI remembers in two completely different ways. Only one of them is ours to control.

Weights

The model's hazy memory

Everything it absorbed in training, crushed into statistics. Lossy, compressed, baked in — and impossible to fully clean.

  • lossy
  • statistical
  • frozen at training time
  • can't be un-baked
Context

The model's exact memory

Whatever we place in front of it right now. Precise, lossless, fully legible — but temporary, gone when the conversation ends.

"Resilience model: stress + recovery balance"framework
"Observational cohort, n=2,400, moderate confidence"evidence
"Holistic lens: sleep, load, social context"framework
  • precise
  • lossless
  • chosen by us
  • ours to control

We can't un-bake the weights. But we control the context. That's the lever.

03 The decision

Three ways to build it. We picked the one in the middle.

Left to right, the cost climbs. The cheapest option lets the bias run the show; the most expensive one buys a different bias at an absurd price. The honest answer sits between them.

cost

Just prompt a stock model

Type a question, take the answer. Fast and nearly free — but the model's baked-in worldview runs the show, invisibly. You get the disease model wearing a friendly tone.

✗ bias runs unchecked
✓ what we built
cost

Ground it on our own curated knowledge base

The model reasons over our vetted material — multiple health frameworks, each with its evidence — and is forced to show its sources. We don't change what it knows; we change what it's looking at.

✓ every claim shows its framework + evidence
cost

Train a model from scratch on "clean" data

Astronomically expensive. Throws away the reasoning ability a big model already has — and still ends up biased toward whatever we decided counted as "clean." Neutral in name only.

✗ costly, and still biased — by us
You don't fix bias by chasing neutrality. You fix it by making every claim show its framework and its evidence.
04 What we built

How it actually works

Your question never goes straight to the model's hazy memory. It passes through the part we control.

User question
"Is my resting heart rate a problem?"
Our pluralistic health knowledge base
Multiple vetted frameworks, side by side
AI reasons over it
Works from our material, not its gut
Answer names its framework + evidence strength
"By the resilience model… (moderate evidence)"
Multiple frameworks, not one truth

A complexity / resilience model sits alongside the conventional biomedical one — held side by side, neither crowned.

Evidence travels with every claim

Each statement carries how strong the evidence is and where it came from — no anonymous assertions.

Disagreement stays visible

When frameworks conflict, we show the conflict. We don't quietly pick a secret winner behind the scenes.

The model's reasoning power stays intact — we just point it at material we trust and make it cite as it goes.

05 The honest part

Grounding helps. It doesn't fully erase the bias. Three ways it leaks through:

01

Interpretation

The model still decides what matters in our text. Trained to trust clinical trials, it can quietly down-rank other valid evidence — even when it's sitting right there in the source.

02

Gap-filling

Where our knowledge base is silent, the model fills the hole with its default — the disease model. Absence in our material becomes presence of the old worldview.

03

Language

Default vocabulary smuggles the worldview back in. Words like "symptom" and "normal range" carry assumptions about what's broken and what's standard.

Notice the color bleeding into these cards. That's the Section 1 bias — still seeping, even through the grounding.

06 How we handle it

We don't claim it's solved. We work it down a ladder — lightest lever first.

Each rung costs more effort than the one above it. We reach for the lightest one that does the job, and only climb when we have to.

1

Grounding rules lightest

Force the AI to attribute every claim to a framework and match its confidence to the strength of the evidence behind it. No naked assertions.

2

Context curation light

Choose what the model sees. Pull in the non-default frameworks deliberately, so the gap-filling has less room to reach for the disease model.

3

Plural prompting moderate

Require more than one lens in the answer. Make the model state the conventional view and the alternatives, then surface where they disagree.

4

Evaluation & red-teaming heavier

Actively probe for the leaks — interpretation, gap-filling, language — and measure how often the default worldview creeps back in. You can't fix what you don't watch.

5

Fine-tuning heaviest · last resort

Only when the cheaper rungs genuinely fall short. Expensive, slow, and it re-bakes our choices into the weights — so we treat it as the lever of last resort, not the first move.

The reframe

We didn't remove the bias. We made it visible, attributable, and contestable.

A "neutral" model is a story you tell yourself until the day it quietly decides what's normal for you. We'd rather show our framework, show our evidence, and let you argue with it. That's the harder path — and the honest one.