AI Ethics
If a system decides who gets a loan or an interview, somebody will eventually ask why. This makes sure you can answer.
What is AI ethics?
AI ethics work turns broad commitments about fairness into specific operational requirements: which decisions must involve a person, what you are able to explain to somebody affected, where historical bias could enter through your own records, and what customers are told about machine involvement.
We keep this concrete. If a system ranks job applicants, screens claims or decides who is served first, then one day a real person will ask why they were treated as they were.
If the honest answer is that nobody knows, that is an operational problem before it is a philosophical one. It is also the point at which a complaint becomes difficult to defend.
Most unfairness enters through your own history rather than through the technology. A system trained on past hiring decisions inherits whatever those decisions contained, including patterns nobody intended.
Why a values statement is not the deliverable
Duara advises on AI ethics from Nairobi, largely for organisations whose systems make decisions about individuals: lenders, insurers, employers, schools and service providers dealing with the public.
We are not interested in producing a values statement. If a commitment does not change what a system does or what a member of staff must check, we will tell you it is decoration and suggest deleting it.
Who this is for
- You use AI in hiring, lending or deciding eligibility
- You serve the public or vulnerable groups
- Customers or funders ask about responsible AI use
- A decision has already been challenged
Industries we serve
- Lending and microfinance
- Insurance underwriting
- Recruitment and HR
- Education admissions
- Healthcare triage support
- Social programmes and NGOs
Problems this solves
These situations tend to arrive together, usually shortly after the first complaint that cannot be answered.
- Decisions nobody can explain. A customer asks why they were declined and the only available answer is that the system decided, which satisfies nobody.
- Bias inherited from your own records. A system built on past decisions reproduces whatever those decisions contained, including patterns nobody intended or noticed.
- No defined human involvement. Consequential decisions run end to end without anybody reviewing them, because no rule required review.
- Customers unaware they are dealing with a machine. Nothing tells people when a decision or conversation is automated, which becomes a trust problem the moment it emerges.
What we deliver
A set of operational requirements covering explainability, human review points, bias testing against your own data, and disclosure to customers, each written as something a system or a person must actually do.
A review of where past data could produce unfair outcomes
An explainability standard for decisions affecting individuals
Defined review points where a person must be involved
Guidance on telling customers when AI is involved
A test method for checking outcomes across groups
A short statement you can publish, backed by the above
Features
- Testing on your own records. Bias is measured against your actual historical outcomes rather than discussed in the abstract.
- Explainability defined as a standard. We specify what must be reconstructable about a decision, not merely that decisions should be explainable.
- Review points tied to consequence. Human involvement is required where the decision materially affects somebody, and not everywhere, which would be unworkable.
- Disclosure wording drafted. The actual sentence a customer sees when a system is involved, written for your context.
- Deletion of decorative commitments. Anything that changes no behaviour is removed, so what remains is enforceable.
Benefits
- Complaints become answerable. When somebody challenges a decision, you can reconstruct it rather than defending a black box.
- Unfair patterns surface early. Testing against your own outcomes finds problems while they are still cheap to correct.
- Funders and clients get a real answer. You can describe what you actually do rather than pointing at a values page.
- Staff know when to intervene. Defined review points remove the ambiguity about whose judgement applies where.
Technologies we use
- Outcome testing across groups. Comparing how decisions fell across categories in your own data to detect disparate results.
- Decision reconstruction. Checking whether your systems record enough to explain a specific past decision months later.
- Review point design. Specifying which decisions stop for a person and what that person is expected to check.
- Disclosure drafting. Writing the wording customers actually see, tested for clarity rather than legal cover.
- Documentation of data lineage. Establishing which historical records fed a system, because that is where inherited bias originates.
The work is analysis and specification. Where testing is needed it runs against your own historical outcomes rather than synthetic data.
How the process works, step by step
Ethics work usually takes three to four weeks and produces requirements rather than a philosophy document.
-
Step 1
Free scoping call
We identify which of your decisions affect individuals and whether any have already been challenged.
-
Step 2
Decision inventory
We list the decisions a system makes or influences, and what each one does to the person on the receiving end.
-
Step 3
Outcome testing
We test past outcomes across groups using your own records to find patterns worth investigating.
-
Step 4
Requirement drafting
We write the explainability, review and disclosure requirements as specific obligations.
-
Step 5
Review with your team
We work through the requirements with the people who will have to meet them, and cut anything unworkable.
Quality assurance and how we work
The commitments attached to ethics work.
- We delete commitments that change nothing. A principle that alters no system behaviour and no staff action is removed rather than published.
- Testing uses your data, not benchmarks. Fairness is measured against your own outcomes, because industry averages tell you nothing about your business.
- We say when a finding is uncomfortable. If your historical data shows a pattern you would rather not see, it goes in the report anyway.
- Requirements are checked for workability. Anything your team cannot actually perform is cut before publication rather than quietly ignored later.
- We do not provide legal cover. This is operational work. Where you need a legal opinion we say so and work alongside your advisers.
Frequently asked questions about AI ethics
What does AI ethics work actually produce?
Operational requirements: which decisions need a person, what must be explainable, how bias is tested against your data, and what customers are told.
Is this the same as compliance?
They overlap but differ. Ethics covers decisions that are lawful but would still damage trust if they became public.
Our historical data is imperfect. Does that disqualify us?
No, most data is imperfect. What matters is knowing where the gaps skew outcomes and deciding deliberately how to handle it.
How do you test for unfairness?
By comparing how decisions fell across groups in your own records. It is measurable rather than a matter of opinion.
What is explainability in practice?
Being able to reconstruct a specific past decision months later, in terms the affected person can follow.
Do we have to tell customers AI is involved?
We recommend it, and we draft the wording. Discovering it later damages trust more than disclosing it upfront ever costs.
Which decisions need a person?
Consequential ones, meaning decisions that materially affect somebody such as credit, employment or eligibility.
Can every decision have human review?
No, and requiring it would be unworkable. That is why review points are tied to consequence rather than applied everywhere.
What if testing shows an uncomfortable pattern?
It goes in the report. A finding you would rather not see is exactly the finding worth having.
Is a published ethics statement enough?
On its own, no. If it changes nothing about what your systems or staff do, it is decoration and we will suggest deleting it.
Who should join the ethics work?
Whoever owns the affected process and whoever handles complaints, because they see the consequences first.
Does this apply if we bought the system rather than built it?
Yes. Decisions made by a tool you purchased are still decisions your business is accountable for.
How long does an ethics review take?
Three to four weeks, most of which is inventory and outcome testing rather than writing.
Do you provide legal advice?
No. This is operational work, and we say clearly where a legal opinion is needed instead.
What is data lineage and why does it matter?
The record of which historical data fed a system. It matters because inherited bias almost always enters there.
Can you review a system already in production?
Yes, and that is common. Live systems have real outcomes to test, which makes the analysis far stronger.
How often should outcomes be retested?
At least annually, and after any change to the data feeding the system or the population it serves.
What if a requirement turns out to be unworkable?
We cut it during the review session. A requirement nobody can meet is worse than no requirement at all.
Does this help with funder or donor questions?
Yes. It lets you describe what you actually do rather than pointing at a values page.
What does this work cost?
It depends on how many decision types are in scope. The figure is agreed before we begin.
Glossary of terms
Terms used in fairness work, defined without the academic framing.
- Explainability
- Being able to reconstruct why a particular decision was reached, in terms a person affected can follow.
- Disparate outcome
- A pattern where decisions fall differently across groups, whether or not anybody intended it.
- Human in the loop
- A requirement that a person reviews a decision before it takes effect.
- Data lineage
- The record of which historical data fed a system, which is where inherited bias usually originates.
- Disclosure
- Telling a customer that a machine is involved in the conversation or the decision.
- Consequential decision
- A decision that materially affects somebody, such as credit, employment or eligibility.
Related services in AI Consulting
Talk to us about AI ethics
If a decision has already been challenged and you struggled to explain it, that specific case is the best place for us to start.
Start your project