Responsible AI principles fail because they're never traced down to the concrete decisions (in data, code, and configuration), where ethics actually gets implemented. This is a map of those decisions. For each stage of a system's life, this matrix shows the question each principle is really asking, what going wrong looks like, and a verified case where it went wrong.
▶ Watch the episodeMost responsible-AI guidance stops at the level of principle: be fair, be transparent, be accountable. These are useful as direction, but doesn't tell you which line of code enacts it. This resource takes the four principles and traces each one to the specific points in a system's life where it stops being a value and becomes a decision someone has to make.
It's built to be read three ways. If you build the system, every live cell is a question for your team. If you buy one, the same cells become questions for the vendor and silence is itself a finding. If you use a system you don't own, the cells tell you what you can and can't see from the outside. The map is deliberately sparse: a cell is filled only where that principle is genuinely decided at that stage. The blank cells are information too.
These four principles recur most consistently across the major reviews of AI ethics guidelines (Jobin et al., 2019; Fjeld et al., 2020; Corrêa et al., 2023). They're a working set and are enough to do real work.
Whether the people affected and the people running it, can see what the system did and why. Provenance, behaviour, and limits documented and disclosed; not a black box presented as a verdict.
Whether benefit and error are distributed justly across groups, and whether the system encodes or amplifies existing discrimination. Rarely a single metric but a choice between incompatible definitions of "fair".
Whether personal data is collected, used, and retained lawfully and proportionately with consent, minimisation, and a defensible basis for every source the system draws on.
Whether there's a named owner, a record of what happened, and a real route to contest and remedy when the system causes harm. The principle that turns "the model decided" back into a human responsibility.
Most responsible-AI material is organised around the traditional ML lifecycle: collect data, preprocess, train, validate, deploy, monitor. That lifecycle assumes the consequential decisions are made before the system ships and are frozen at deployment. That assumption breaks the moment you leave traditional ML.
A RAG system often has no training step at all; its behaviour changes when you edit a retrieval corpus, after deployment, with no retrain and usually no evaluation gate. An agent makes its most consequential choices at runtime (which tool to call & whether to act) continuously, not once. So this resource replaces the named lifecycle stages with six functional stages that describe what every system does, regardless of type. Predictive systems concentrate their ethics decisions early; RAG pushes them into a live corpus; agents push them into runtime.
| Problem definition / scoping | → | Framing |
| Data collection & labelling | → | Foundations |
| Model training & feature engineering | → | Behaviour |
| Testing & validation | → | Evaluation |
| Deployment & inference | → | Runtime |
| Monitoring & maintenance | → | Monitoring |
Pick a system type, then click on any lit cell within the matrix to open it.
This resource is shared to do three things: pass on what I've learned, invite feedback from others working in this space, and contribute to the wider conversation about building better norms and practices around responsible AI.
If you've used it, adapted it, or have thoughts on the framing, I'd love to hear from you. Reach me through the contact form or on LinkedIn.