AI ethics and bias

Three ways to see it

1. 1

   Bias in AI enters through three doors. The first is data. If the past records you train on are themselves biased, the model will pick the bias up as a pattern. The second is labelling. Humans tag the examples; their tagging carries their assumptions. The third is deployment. A model trained for one purpose, used in another, can hurt people the original designers never met. All three matter. None of them goes away by waiting.

2. 2

   Pakistan adds a particular layer. Most foundation models are trained on Western-English data. They do better on a CV from a Beaconhouse School graduate than on a CV from a Government High School graduate in Larkana, even when the second person is more qualified. They do better on Urdu in Nastaliq than on Sindhi in Arabic-Sindhi script. They do better on questions about Texas property law than on questions about Punjab Tenancy Act. None of this is a moral judgement of the model. It is a description of which examples it has seen the most.

3. 3

   Pakistan's legal scaffolding is taking shape. The Prevention of Electronic Crimes Act (PECA), the Personal Data Protection Act 2023, and the draft National AI Policy from the Ministry of IT and Telecom together start to define what a responsible AI deployment looks like, what consent for data use means, and what redress a citizen has when an algorithm gets them wrong. These are not finished frameworks. They are the early outline of one. A working professional should at least know they exist.

The why-tree

Try this with Claude

AI-edge prompt: 'I'm deploying an AI tool to do [task] in a Pakistani [sector] context. List five ways this deployment could produce unfair or harmful outcomes for specific groups, with one mitigation each. Be concrete; use Pakistani examples.' Treat the answer as a starting checklist, not a verdict.