Responsible AI in African markets: what good governance looks like
Most organisations that publish AI principles have not changed anything as a result. That is not evidence of bad faith. It is what happens when a commitment has no owner, no budget and no process attached to it.
Governance is an operating question
Useful governance looks boring: a named owner for each system, a documented risk classification, a defined human review path, and a record of what decisions the system has influenced. None of that requires a policy department, but it does require someone accountable.
Context changes the risk
A model applied to product recommendations carries different risk than the same model applied to credit decisions, and neither is meaningfully comparable to its use in medical triage. Risk classification has to happen per deployment, per market, because the consequences and the available remedies differ.
- Who can override the output, and are they trained to do so?
- What data was used, and would you be comfortable publishing that list?
- What happens to someone who is harmed by a decision, and how would they know?
If a person cannot explain, in plain language, what an automated system decided about them, the system is not finished.
Regulation will arrive unevenly
Enforcement capacity, not legislation, will determine how these systems are actually built. Organisations that build documentation now will be the ones able to answer a regulator's questions in an afternoon rather than a quarter.
