Ethical AI Framework
An ethical AI framework is the set of checks you run before and after deploying a model, who's accountable, what could go wrong for real people, and how you'd know, so ethics isn't decided ad hoc by whoever shipped the feature.
A line of checkpoints runs either side of deployment, each one a question someone has to answer by name.
Reach for this when…
- You're about to put a model in front of customers and nobody owns the risk.
- An AI policy document exists but nobody consults it before launch.
- A model behaved badly in production and no one can say who signs off on the next one.
How to run it
- Name who is accountable for each AI system's outcomes, by name, not by department.
- List the ways it could cause real harm: wrong decisions, bias, leaked data, unclear reasoning.
- Test for bias and error rate across the groups it actually affects, before launch.
- Decide what gets explained to the person affected, and build that in.
- Monitor after launch and set the trigger for pulling or retraining it.
A worked example
Situation. Lukas Hoffmann runs MediKlar, a small health-tech SaaS company in Leipzig, Germany, that had built a triage chatbot no one had formally reviewed for how it handled uncertain symptoms.
Applied. Before wider rollout he named himself accountable, tested the bot's advice against a set of ambiguous cases, and built in a hard rule that anything uncertain routes to a nurse, with the reasoning shown to the patient.
Result. The rollout went ahead with a documented owner and a routing rule that caught the cases the bot was worst at, instead of shipping first and finding out from a complaint.
The catch
There is no single agreed ethical AI framework; most organisations assemble their own from overlapping principles, which means the checklist can become theatre if nobody has real authority to stop a launch. Bias testing also only catches what you thought to measure.
A framework with no one empowered to say no before launch is a values statement, not a control.