Kobalt Research

Home · Applications · Climate & Biosphere

Climate AI must include
the biosphere as an entity

Ecosystems have welfare functions with tipping points — non-linear thresholds after which restoration is impossible. Future generations have welfare functions that standard exponential discounting reduces to zero. Indigenous communities have welfare knowledge that no climate model has formalized.

Application · Climate Implications for Climate & Biosphere Focus: Google.org Impact Challenge 2026
— The structural problem

Climate AI systems optimize for aggregate outcomes (global temperature, GDP loss, emission reductions) while the welfare consequences concentrate on specific entities: island nations, indigenous communities, Arctic ecosystems, future generations. The disaggregation cost is borne by exactly those entities with the least political weight in the optimization.

— Entities excluded from climate AI

Who the current framework misses

Ecosystems with tipping points
Non-linear welfare functions

Coral reefs, permafrost, Amazon rainforest — all have non-linear welfare functions. Below certain thresholds: high welfare. Above: catastrophic, irreversible loss. Standard linear models cannot represent this.

Future generations
Discounting destroys representation

Every year of delayed climate action shifts welfare costs forward. Under standard exponential discounting ($\delta=0.03$), generations in 2124 are worth 5% of today. Under $w(T) = 1 + \ln(T)$: 5.6×. The mathematical choice determines the policy conclusion.

Indigenous communities
Welfare annihilation, not degradation

Communities whose entire welfare function is embedded in ecosystem health — diet, medicine, spiritual practice, territorial identity. When the ecosystem collapses, these welfare functions are annihilated, not degraded.

Small island developing states
Binary welfare discontinuity

Sea-level rise creates a binary welfare function: the island exists and welfare is defined; the island is submerged and welfare is undefined. No climate model represents this discontinuity formally.

— Tipping points as welfare functions

Why linearity fails

The welfare function of a coral reef is not linear. It looks like: $W_e(\text{temperature})$ is constant for $T < 28°C$, then drops catastrophically for $T > 28°C$. This is not modeled by any standard climate-economic model.

$$W_e^{\text{ecosystem}}(T) = \begin{cases} W_0 & T < T_{\text{crit}} \\ W_0 \cdot e^{-\lambda(T - T_{\text{crit}})} & T \geq T_{\text{crit}} \end{cases}$$

— Ecosystem welfare function with non-linear tipping point at $T_{\text{crit}}$

Every global climate model aggregates ecosystem welfare into GDP projections, hiding this non-linearity behind averages. The Jensen gap — the difference between the welfare of the average temperature and the average welfare across temperature distributions — is not a rounding error. It is the measure of how much is being concealed.

— Indigenous knowledge as formal input

The PCC-W application

The PCC-W protocol for climate entities: local ecological knowledge (phenology, species behavior, water patterns) is formal input to $W_e$ construction, not qualitative context. Communities that have observed an ecosystem for generations have data that no sensor network can replicate.

— Protocol note

A community that has tracked glacier retreat for 60 years has 60 years of welfare-relevant data about the ecosystem. The protocol formalizes how to weight this against satellite measurements, climate model outputs, and biodiversity indices in the construction of $W_e$.

This is not a concession to qualitative methods. It is a recognition that temporal depth of observation is a formal epistemic input — one that sensor-based monitoring systems began accumulating only recently, and that communities have been accumulating for generations. The protocol specifies how to combine these inputs without subordinating either.

— Conceptual dependency map

This page depends on

Pages that extend this one