---
kind: concept
slug: life-cycle-assessment
title: Life cycle assessment for AI services
summary: >-
  Judging an AI service by its whole life cycle, from raw materials through production, use and disposal, rather than by training or use-phase energy alone. The method compares an AI-enhanced application with a reference application, impact by impact. One position paper proposes it, with no experiments or code, so the page has a single source.
tags: [sustainability, life-cycle-assessment, environmental-impact]
papers: ["2110.11822"]
related: [carbon-accounting, carbon-aware-scheduling]
status: draft
updated_at: 2026-11-08
---

## Overview

Life cycle assessment (LCA) adds up the environmental flows of a product or
service across its stages: raw material acquisition, production, use and end
of life [arxiv:2110.11822 §3.2, tab. 1]. The question it answers for AI is
whether an AI-enhanced application has a smaller environmental footprint than
the same application without AI, once the AI's own impacts are counted
[arxiv:2110.11822 §4.1, eq. 1].

The shared idea is a difference. Each system is assessed over the same
functional unit, and the result is the change in each impact criterion. A
positive change means the AI-enhanced application causes more impact
[arxiv:2110.11822 eq. 1]. The paper does not state this sign convention in
words; it follows from how the difference is defined.

## Variants

### Paired comparison with a reference application

This is the method the paper proposes. Two systems are assessed: a reference
application without AI, and the AI-enhanced application, whose assessment
includes the AI service [arxiv:2110.11822 §4]. The per-criterion difference
is the net impact [arxiv:2110.11822 §4.1, eq. 1]. The paper's worked example
is heating one square metre to 20 °C for a year [arxiv:2110.11822 §2.3]. The
AI-enhanced case it discusses is a building whose AI saves heating energy by
tracking occupancy [arxiv:2110.11822 §4.1]. Card:
[Unraveling the Hidden Environmental Impacts of AI Solutions for Environment](/papers/2110.11822).

### Consequential assessment at large scale

The paper says that a large deployment needs a consequential LCA, which also
counts knock-on effects such as new mines or extra fossil-fuel generation
[arxiv:2110.11822 §5.4]. The source card records only this statement, with no
procedure or numbers for it, so treat it as an open direction rather than a
method [arxiv:2110.11822 §5.4].

## When to use which

| Situation | Use | Why |
|---|---|---|
| You can name a reference application that does the same job without AI | Paired comparison | Only the difference shows whether the AI gives a net gain [arxiv:2110.11822 §4.1, eq. 1] |
| You have only the energy of training or running the model | Do not report a net gain | The paper says energy alone misses production and end-of-life impacts [arxiv:2110.11822 §2.1, §3.1] |
| The deployment is large enough to change supply chains or energy markets | Consequential assessment, as a direction | The attributional comparison may miss knock-on effects [arxiv:2110.11822 §5.4] |
| You need one score to rank options | Keep impact criteria separate | The paper argues against collapsing criteria into one number and rejects monetary value units [arxiv:2110.11822 §4.1] |
| You want the benefit to be stated in a sales or policy claim | Show the full comparison and its inputs | The paper says a sound LCA shows technical potential only; real-world effects can differ [arxiv:2110.11822 §5.3] |

## Implementation sketch

The steps follow the paper's method [arxiv:2110.11822 §3.2, §4]. Impact
criteria and the functional unit are inputs the user must supply.

```pseudocode
# Net life-cycle impact of an AI-enhanced application  [arxiv:2110.11822 eq. 1]
# Input:  reference app m1, AI-enhanced app m2, d impact criteria
# Output: delta, one value per criterion; positive means more impact

def net_impact(m1, m2, criteria):
    # Assess both systems over the same functional unit  [§2.3]
    # Scope each tier: terminals, network, data centre or server  [§3.2]
    lca_m1 = life_cycle_assessment(m1, criteria)
    lca_m2 = life_cycle_assessment(m2, criteria)
    # The assessment of m2 includes the AI service's own impacts  [§4.1]
    # Keep one value per criterion; do not collapse them  [§4.1]
    return lca_m2 - lca_m1
```

Shared equipment needs an allocation rule. The paper suggests execution time
for production and end-of-life of shared servers, and says the exact key is not
stated [arxiv:2110.11822 §3.2]. The static use-phase share in its example is
1/n for n concurrent programs [arxiv:2110.11822 §3.2].

```pseudocode
# Share of a device's static use-phase energy charged to one program  [§3.2]
# Input:  static_energy (J), n_concurrent (int, >= 1)
# Output: energy charged to this program (J)

def static_share(static_energy, n_concurrent):
    # Split evenly: each of n concurrent programs takes 1/n  [§3.2]
    return static_energy / n_concurrent
```

## Contradictions

None known. The page rests on one paper, and that paper reports no experiments
whose results could conflict with another's.

## Papers

- [Unraveling the Hidden Environmental Impacts of AI Solutions for Environment](/papers/2110.11822): proposes a life-cycle comparison of an AI-enhanced application with a reference application, and reviews AI-for-Green papers for environmental evaluation.