The pillars of trustworthy AI, operationalised

Three ways to see it

1. 1

   Accountability means a named person carries the can. Not the model. Not the vendor. A human inside your organisation must own each AI system end to end. Transparency means the affected person can find out what was decided about them and what kind of system did it. Fairness means the system does not produce systematically worse outcomes for protected groups without justification. Robustness means the system behaves reasonably under unusual inputs and adversarial attempts. Privacy means data is collected, used, and retained narrowly and with consent. Sustainability means the system's lifecycle impact, energy, water, and computing waste, is tracked and reduced where possible.

2. 2

   Way one to operationalise: name owners. For each pillar, write a name and a role. Accountability owner often sits with the project sponsor or chief risk officer. Transparency owner usually sits with product or communications. Fairness owner often sits with data science but escalates to legal. Privacy owner is the Data Protection Officer once PDPA passes. Robustness owner is engineering. Sustainability owner is whoever runs ESG or facilities. Without names, every pillar is owned by no one.

3. 3

   Way two: write tests. A pillar that cannot be tested is not a pillar, it is a slogan. Fairness can be tested by disparate impact ratios across protected groups; in Pakistan you would test across province, gender, language, and economic band. Robustness can be tested by adversarial inputs, perturbed inputs, and stress conditions. Privacy can be tested by re-identification attempts on supposedly anonymised data. Transparency can be tested by asking five random users to explain in their own words what a system did to them.

The why-tree

Try this with Claude

AI-edge prompt to try: 'Acting as a trustworthy AI auditor, propose three concrete fairness tests appropriate for a Pakistani credit scoring model. For each test name the protected axis, the metric, and the threshold at which I should flag a problem.' Use the response to prime a discussion with your data team, not to settle it.