Manage the Trade-offs Between the Dimensions Deliberately
When protecting one thing weakens another, who decides the balance — and on what basis?
Enhanced: 1Advanced: 12 controls
Who it applies to
Enter for every AI system whose letter is E or A (this area has no Baseline controls). Go as deep as your system letter, and implement in your mode.
Why it exists
Recognise that the things this framework protects sometimes pull against each other, and manage the tension as an explicit decision rather than an accident of whoever built the system. The dimensions are interdependent but not always aligned: hardening a model against attack can make it less accurate; protecting privacy strongly can reduce a system’s usefulness or its fairness; logging everything for accountability creates a privacy exposure and a cost. You cannot maximise all of them at once, and pretending otherwise means the trade-offs get made silently, by default, by the person closest to the code. Make them instead a deliberate decision, weighed against your risk appetite, sector, and the specific use — and this is the domain where that decision lives.
What to actually do
Surface the conflicts between the dimensions, and decide them on purpose. When strengthening one protection weakens another, name the conflict, weigh it against the organisation’s risk appetite and the stakes of the specific use, and record the decision and its reasoning. The goal is not to eliminate the tension — it cannot be eliminated — but to ensure it is resolved by someone accountable, on stated grounds, rather than by inertia.
Let proportionality settle the balance, not uniform maximisation. Apply the same proportionality that runs through the whole framework: how far you push each dimension depends on the role, the sector, the criticality, and the size — so the balance struck for a small consumer’s low-stakes tool differs, legitimately, from the balance struck for a regulated developer’s high-risk model.
Controls
Enhanced1 controls
Control 231
Resolve conflicts between the dimensions as explicit, recorded decisions.
EnhancedPreventiveall rolesper system
What it is
Weighed against risk appetite and use, owned by an accountable person.
How to implement it · Minimum content
Conflict: the dimensions pulling against each other (e.g. attack hardening vs accuracy; privacy vs fairness testing; full logging vs privacy and cost)
Options considered and the effect of each on every dimension involved
Context: risk appetite, stakes of the use, people affected
Decision taken and its reasoning
Accountable person who decided, and date
Review trigger: e.g. change of use, model or applicable rules
How to start
Ask each system owner which trade-offs have already been made by default (e.g. log retention, filter strictness) and record the most significant one.
Evidence
The trade-off decision log with reasoning and the accountable decision-maker for each entry.
Mode: light / standard / full
light
The owner-manager decides and notes each trade-off in the system's file.
standard
Owners log trade-offs; those touching people or high risk go to management.
full
Trade-offs go to the AI committee as formal, logged decisions.
Advanced1 controls
Control 232
Set the balance by proportionality, not uniform maximisation.
AdvancedPreventiveall rolesper system
What it is
Calibrated to role, sector, criticality, and size.
How to implement it · Minimum rules
Calibrate each dimension to four factors: role, sector, criticality of the use, organisation size
Write a calibration table: target settings per criticality level (e.g. logging depth, filter strictness, testing frequency)
Do not default to the maximum on every dimension; record what each setting costs in the others
The Baseline is the floor in every dimension, whatever the calibration
Recalibrate when role, sector, criticality or size changes
Compare systems periodically: settings should differ where context differs
How to start
Compare the balance struck in two systems of different criticality and write down why it differs, or why it should.
Evidence
The calibration table and per-system settings showing the balance varies with context.
Mode: light / standard / full
light
Management approves a half-page settings table applied to the few systems in use.
standard
The risk lead keeps the table and checks each system against it at approval.
full
The AI committee owns the standard; second-line review checks for uniform maximisation.
Evidence
The recorded trade-off decisions and their reasoning; evidence the balance is calibrated to context rather than uniformly applied.