Ai ethics is currently a declaration sport. Organisations publish principles, frameworks, and commitments, and then have no structured way to demonstrate whether they are honouring them.
Procurement teams cannot compare ethical standards across vendors. Boards cannot hold products accountable. Regulators are writing legislation faster than the industry is developing the means to comply with it.
This framework exists to close that gap. It converts principles into measurements, measurements into scores, and scores into something you can act on, report on, and be held to.
Everything in this framework traces back to four principles. They are not aspirations. They are the dimensions along which every system is assessed, scored, and reported on.
Consent
Data is collected and used with consideration for ownership, permission, and legal provenance. This includes how data was gathered, whether individuals consented to its use in Ai training, and whether removal or correction requests can be honoured.
Was the data used responsibly?
Representation
A system is only as fair as the perspectives it was trained on. Representation measures source diversity, geographic balance, language coverage, and whether dominant narratives are being amplified at the expense of others.
Whose voices are represented?
Provenance
Every input and output should be traceable to credible, documented sources. Provenance covers explainability of decisions, attribution of outputs, and how the system handles historic content that no longer reflects current standards.
Can we explain how it thinks?
Accountability
Decisions must be explainable, auditable, and monitored over time. Accountability covers governance structures, misuse controls, drift detection, and what happens when a system's behaviour diverges from its stated values.
Will it stay ethical over time?
Before you commission, procure, or sign off on an Ai system, these are the five questions that need answers. Not asked once in a scoping session and forgotten, but built into your supplier requirements and revisited at every major stage of the project.
Does the data reflect diverse global perspectives, not just dominant narratives?
Was it collected with consideration for consent, ownership, and legal provenance?
Have we measured and mitigated bias in both training data and outputs?
Are the Ai's decisions traceable to credible sources with clear reasoning?
Do we have safeguards to monitor model drift and unintended consequences?
Principles without measurement are intentions. The scoring system converts the four principles into weighted, auditable dimensions, each with defined indicators, numeric outputs, and hard floors that disqualify rather than penalise.
A single composite score. Four dimension scores. A public audit trail. Something you can put in a procurement response, a board paper, or a regulatory submission.
| Dimension | Weight | What it measures |
|---|---|---|
| 30% | Data provenance, licensing, collection practices, deletion compliance | |
| Representation | 30% | Source diversity, geographic balance, language coverage, amplification ratio |
| Provenance | 20% | Output traceability, explainability level, source credibility, dataset age |
| Accountability | 20% | Drift detection, misuse incident rate, human oversight coverage, audit cadence |
These are not low scores, they are disqualifications. No composite score is issued if any floor is breached.
Two instruments. One positions your Ai system in the space between declaration and demonstration. The other reveals the shape of the risk beneath the score. Together they form the Nexus Ai™ Ethics System — a proprietary diagnostic method, not a scorecard.
Compliant but fragile
Data is well-governed but behaviour diverges. Ethics declared, not embedded.
✓ Verified
Data integrity and behavioural integrity align. Ethics is demonstrable.
Ethical theatre
Ethics exists as policy only. Neither the data nor the outputs can be defended.
Lucky
Behaviour appears ethical now, but the data foundations won’t hold under scrutiny.
The Matrix and the Profile are designed to be read together. The Matrix tells you where your organisation sits in the space between declaration and demonstration. The Profile tells you the shape of why — and where the specific failure will come from. A system in the Verified quadrant with a uniform profile is genuinely strong. A system in the Verified quadrant with a 35-point gap between its best and worst dimension is one audit away from a reclassification.
Nexus Ai™ · Ethics System v1.0
Ethical risk rarely announces itself. These ten domains map where it hides, each one tagged to its core principle and actively monitored throughout the system's life. Naming risk is the first step to managing it.
Can users be identified through combined metadata, even when data appears anonymised?
Do time patterns in data risk misrepresenting intent, agenda, or evolving context?
Are dominant geopolitical perspectives being reinforced at the expense of regional voices?
Could combined metadata enable re-identification in ways individuals did not consent to?
Could pattern signatures be misused to monitor, target, or suppress individuals or groups?
What are the consequences of linking sentiment or engagement scores back to specific contributors?
Does model training inadvertently amplify the loudest voices while silencing minority perspectives?
How might outputs or data assets be exploited by bad actors for influence, manipulation, or control?
Is it ethical to score or rank contributors by inferred impact? What harms does that enable?
How do we handle historic content that no longer reflects current ethical norms or factual standards?
Ethics isn't a checklist you run at the end of development. It's a continuous practice, embedded at each stage of the Ai system's life, from the moment data is ingested to the point at which the model is retired.
Consent checks, provenance verification, diversity assessment
Bias detection, source balancing, historic content flagging
Fairness thresholds, explainability requirements, concentration limits
Measurable indicator scoring, red-teaming, floor checks
Access controls, misuse monitoring, human oversight protocols
Drift detection, scheduled re-evaluation, third-party audit cadence
The framework is the methodology. What we provide is the expert practice of applying it: structured engagement, independent assessment, and outputs that hold up in a procurement process, a boardroom, or a regulatory review. This is not a self-assessment tool. It is a managed service, delivered by practitioners who built it.
A structured assessment of your Ai system against the full framework. Per-dimension scores, floor status, active risk flags, and a prioritised improvement roadmap. Formatted for procurement responses, board papers, and regulatory submissions.
A scored, defensible ethics record.
Bronze / Silver / Gold / Verified certification tiers based on composite score and floor status. A public badge tied to a documented methodology, usable in pitches, on product pages, and in tender responses.
A credential that means something.
Live score tracking across all four dimensions. Drift alerts, floor warnings, and scheduled re-evaluation triggers. Delivered as a managed service or self-hosted dashboard with quarterly reporting.
Ethics that doesn't lapse between audits.
Quarterly audits, proactive risk identification, score improvement recommendations, and board-level reporting support. Positions ethics as an ongoing strategic capability, not a compliance exercise.
Ethics as a competitive advantage.
A team at Wits University, working with global climate journalism non-profits ClimateXchange and Syli, wanted to build an LLM-based tool capable of analysing four decades of climate news reporting for trends, narrative triggers, and the factors that make coverage actually inspire action.
Their starting dataset contained more than seven million records, pulled from a crawl of news sites and archives spanning 1968 to 2025, across 35,000+ outlets worldwide. It was large, ambitious, and, from an ethical standpoint, entirely unexamined.
Applying the Nexus Ai™ framework forced the team to assess each source against the four principles. Questions of consent, representation, provenance, and accountability eliminated millions of records: scraped without permission, geographically imbalanced, unattributable, or too old to reflect current ethical norms.
Ethics by design is not just a filter for removing what is wrong. It is a lens for amplifying what is right. The smaller, ethically validated dataset produced more objective outputs, surfaced more representative climate narratives, and made the tool more useful to the journalists it was built for. The process of applying the framework did not constrain the project's ambition. It is what made that ambition achievable.
Maai, the tool that emerged from this process, is now positioned as a climate narrative intelligence platform for newsrooms worldwide, built on a foundation that can be demonstrated, audited, and defended. One tenth of the original dataset. Significantly better performance. A product that holds up to scrutiny because the ethics work happened by design, not as a retrospective filter.