Methodology · Taxonomy

The TriAxis Taxonomy

Each bias is classified according to 3 orthogonal axes that answer 3 simple questions: Where does the bias come from, at what stage does it occur, and what is its specific nature.

Pourquoi TriAxis

A classification that stands the test of time

Most existing taxonomies classify biases by symptom (what the bias does) or by domain (health, justice, recruitment...). These approaches are useful but they age poorly: new biases constantly appear in new contexts.

TriAxis adopts a structural approach: it describes each bias by 3 fundamental properties that depend neither on the domain nor on the technology. A bias remains classifiable in the same way whether it is detected in 2020 or 2035.

O

Origin

Where does the bias come from? 3 possibilités : Systemic, computational, or human-cognitive.

L

Lifecycle

At what stage of the AI Lifecycle? 6 stages from L1 (Historical) to L6 (Deployment).

S

Specific

What is the specific nature? 30 categories from S01 (proxy) to S30 (interaction).

Les 3 Axes

Details of Axes O, L, and S

O

Origin Axis

Where does the bias come from?

O1

Systemic

The bias is embedded in social, historical, or institutional structures. It precedes AI and is reflected in its data.

O2

computational

The bias arises from the algorithm itself: model choice, optimization, data representation, feedback loops.

O3

Humain-cognitive

The bias comes from the cognitive biases of designers, annotators, or users: anchoring, confirmation, availability.

L

Lifecycle Axis

À quelle étape du Lifecycle IA?

L1

Historical

Historical data, archives, past corpora carrying traces of the past.

L2

Representation

Choice of variables, encoding, tokenization, representation vectors.

L3

Measurement

Metrics, thresholds, labels. What is measured determines what the model learns.

L4

Aggregation

Data fusion, weighting, averages that perpetuate disparities.

L5

Evaluation

Choice of test sets, performance metrics, acceptance thresholds.

L6

Deployment

Production release, usage drift, feedback loops, progressive degradation.

S

Specific Axis

Quelle est la nature précise du biais?

30 categories covering all bias mechanisms identified in the literature. EXAMPLES:

S01
Proxy
S02
Feedback
S03
Selection
S04
Sampling
S05
Annotation
S07
Distribution shift
S10
Automation
S24
Institutional

Complete list in the database (30 categories).

SCORING SYSTEM

The IBI Score

Each bias receives an IBI score (0 à 10) calculated from 4 quantitative dimensions : Severity, Frequency, magnitude of impacted groups, and difficulty of Detection.

⚠️

Severity

From 1 (minor) to 5 (critical). Potential impact on individuals.

🔁

Frequency

Rare, occasional, frequent, systematic.

👥

Group magnitude

Number of impacted protected groups (1 to 5).

🔍

Detection

Difficulty to identify the bias (1 = easy, 5 = very difficult).

Severity scale (Official levels)

5 - Critical
Direct risk to fundamental rights. Deployment to be prohibited or restricted.
4 - Severe
Significant discrimination identified. Mandatory correction.
3 - Moderate
Measurable but limited discrepancy. Monitoring and documentation.
1 to 2 - Low
Marginal impact. Continuous improvement.

Exemple

TriAxis in action

Here is how a classic bias is described with TriAxis.

IBI-0002Score 9.0Critical

Discriminatory proxy bias

Apparently neutral variable serving as a statistical substitute for a protected attribute.

Origin
O2_computational

Rooted in the algorithm and variable selection.

Lifecycle
L3_Measurement

Manifests in the choice of metrics.

Specific
S01_proxy

Substitution by correlated variable.

Explore the database

250 biases, 68 Regulations, 21 analytical views. Everything is searchable, citable, and open.