Ethical AI is often discussed as a set of values: fairness, transparency, accountability, privacy. The values matter, but on their own they do not change how a system behaves. What changes behaviour is engineering: the decisions built into how a system is designed, evaluated, deployed and corrected.
This guide sets out the practices we treat as requirements rather than extras, and how each one shows up in real systems.
Fairness: measure it, do not assume it
A model trained on historical data can reproduce the patterns in that data, including unfair ones. It can also perform well on average while performing badly for a particular group. You cannot tell either by looking at the model.
Fairness therefore has to be measured. In practice:
- Identify the groups your system affects, and which differences in outcome would be unacceptable.
- Break results down by those groups. An overall accuracy figure can hide serious gaps.
- Look before users do, so problems are found in testing rather than reported by the people harmed.
Transparency: make decisions traceable
When a system makes or supports a decision, the people affected, and those accountable for it, should be able to understand why. That is very hard to add after the fact, so it has to be designed in from the start.
In practice it means logging what the system saw and what it produced, and, for systems that answer questions, citing sources. In Nexo, every answer comes with auditable citations back to the documents it relied on. In a compliance setting, an answer you cannot trace is an answer you cannot use, however likely it is to be correct.
Accountability: keep a person where it matters
Automation should not remove responsibility. Every system needs a clear line between the decisions it may make alone and those that must reach a person, and that line should be designed deliberately rather than left to chance.
Two examples from our own work:
- Our Advanced NER Pipeline routes uncertain extractions to human review, and feeds corrections back into the model.
- Megan routes anything needing clinical judgement to a clinician, so the system handles routine questions and nothing more.
In both, escalation is a designed path rather than a fallback for failure.
Privacy: use less real data
The most reliable way to protect personal data is not to collect or expose it in the first place. Minimise what you gather, remove identifiers the system does not need, and keep data only as long as it serves its purpose.
Synthetic data is one practical route. Our synthetic survey work produces research insight without relying on a live respondent panel. Where synthetic data can do the job, it removes a category of risk entirely.
Reliability: know when the system is unsure
An often overlooked part of ethical AI is honesty about uncertainty. A system that is confidently wrong causes more harm than one that says it does not know. Good systems estimate their own confidence, and route low-confidence cases to people or decline to answer.
Evaluation: before, during and after
An ethical review at launch is not enough. Systems change as they are updated, and the world they operate in changes around them. Evaluation should run:
- Before release, against representative data and the groups the system affects.
- In production, monitoring outcomes, not only uptime.
- After every change to the model, its data or its prompts.
Security is part of ethics
Commitments about fairness and privacy mean little if a system can be misused or its data taken. Access controls, protection against manipulation of inputs, and secure handling of data are ethical requirements as much as technical ones.
A checklist
- Have you measured performance across the groups this system affects?
- Can you trace any output back to what produced it?
- Is it clear which decisions reach a person, and does that actually happen?
- Does the system recognise and act on its own uncertainty?
- Are you using the least personal data you need?
- Does evaluation continue after launch?
If the answer to any of these is no, that is the next piece of engineering work. Ethics in AI is not a separate discipline; it is a standard of engineering.