Responsible AI in Practice
If a commitment cannot be broken, it is not doing anything.
What is responsible AI?
Responsible AI in practice means turning stated principles into specific rules that can be checked and broken, covering which decisions require a person, what an affected individual can be told, and how outcomes are tested for unfairness.
Most responsible AI statements share one property. They cannot be violated. A commitment to fairness and transparency has no state of the world that breaks it, which is why organisations can publish one and change nothing.
A usable commitment is specific enough to fail. No application is declined without a person reviewing the file. Anybody refused can be told the main reason within a working day. Both of those can be broken, which is what makes them worth writing.
The second thing that goes wrong is the human in the loop. A reviewer who approves everything the system suggests is not oversight. That arrangement automates the decision and adds somebody to blame for it, which is worse than automating it honestly.
Why we do not write principles
Duara works on the operational side of responsible AI in Kenya, turning stated principles into rules that can be checked.
We do not write principles. Plenty of people will draft you a values statement, and it will change nothing about what your systems do, which is the reason your existing one has not.
Who this is for
- You have published AI principles with nothing behind them
- Your systems make decisions affecting people
- Customers or funders ask about responsible use
- You cannot explain a decision to somebody it affected
Industries we serve
- Banking and lending
- Insurance
- Healthcare
- Recruitment
- Public services
- Any organisation making decisions about people
Problems this solves
These are the specific ways responsible AI intentions fail to reach the system.
- Commitments that cannot be broken. A principle with no failing case changes nothing, which is why publishing one is so easy and so common.
- Oversight that always agrees. A reviewer approving every recommendation is a rubber stamp, and it converts a person into somebody to blame rather than somebody deciding.
- Explanations aimed at the wrong person. Describing the factors a model weighed is not an explanation to somebody who was refused and wants to know what to do next.
- Bias never actually measured. Unfair outcomes are assumed absent because nobody intended them, and intention is not a property of data.
What we deliver
Commitments specific enough to break, oversight measured by how often it overrides, explanations tested on the people receiving them, and outcomes checked across groups in your own data.
Commitments written so they can be violated
Human review defined at the decisions that matter
Override rates measured rather than assumed
Explanations an affected person can act on
Outcome checks across groups in your own data
A written account of what we found, including the awkward parts
Features
- Commitments with failing cases. Each one written so there is a state of the world that breaks it, since that is the only kind worth publishing.
- Oversight measured by overrides. A reviewer who never disagrees is not providing oversight, and the override rate is the number that shows it.
- Explanations for the affected person. Written so somebody refused understands the main reason and what would change it, rather than how the model works.
- Outcomes checked in your data. Testing results across groups, because a system trained on past decisions reproduces past decisions including the indefensible ones.
- Findings reported as found. Including results you would rather not have, since a check that only confirms what you hoped is not a check.
Benefits
- Principles that change something. Specific rules alter what systems and staff do, which values statements have never managed.
- Real oversight rather than cover. A reviewer with genuine authority protects the organisation, and a rubber stamp exposes an individual instead.
- Fewer complaints escalating. People who receive a comprehensible reason argue less than people who receive a decision without one.
- Problems found by you first. Unfair outcomes discovered internally are a fixable issue, and the same outcomes discovered publicly are not.
What responsible use actually requires
- Testable commitments. Rules written with a failing case, so compliance is a question of fact rather than of opinion.
- Human review that bites. Defined at the decisions that matter, with authority to overturn and a measured override rate.
- Outcome testing. Checking results across groups in your own data rather than assuming fairness from intention.
- Explainability. What an affected person can be told, tested on whether they can act on it.
- Appeal routes. How somebody challenges a decision, which is the part most systems omit entirely.
- Recorded decisions. Keeping enough about each decision that it can be reconstructed later, since complaints arrive after the fact.
Bias is a property of data rather than of intention. A system trained on your past decisions will reproduce your past decisions, including any you would struggle to defend, and no amount of good faith in the design affects that.
How the process works, step by step
This work usually takes three to six weeks, and the outcome testing depends on how much decision history you hold.
-
Step 1
Free review call
We look at what your systems decide about people and what you have already committed to publicly.
-
Step 2
Commitments made breakable
Each principle rewritten as something specific enough to have a failing case.
-
Step 3
Oversight points defined
Which decisions need a person, with the authority to overturn rather than to sign.
-
Step 4
Outcomes tested
Results checked across groups in your own data, and reported as found.
-
Step 5
Explanations written and tested
Checked on whether an affected person understands them, which is a different test from whether they are accurate.
Quality assurance and how we work
The commitments attached to work that can produce unwelcome findings.
- We do not write values statements. A principle that changes nothing about your systems is not worth paying for, and that is what most of them are.
- We report what the testing finds. Including outcomes you would rather not have, because a check that can only confirm what you hoped is not a check.
- We say when oversight is decorative. A reviewer approving everything is not oversight, whatever the process document says about it.
- We say when a system should not be built. Some decisions should not be made this way, and that answer is occasionally the right one.
- We test explanations on people, not on ourselves. An explanation that satisfies the team that wrote it tells you nothing about the person receiving it.
Frequently asked questions about responsible AI
What is wrong with our AI principles?
Probably that they cannot be violated. A commitment with no failing case has no effect on what your systems or staff do.
What does a usable commitment look like?
Something specific enough to fail. No application is declined without a person reviewing the file is a commitment, because it can be broken.
Is human oversight enough?
Only if the human sometimes disagrees. A reviewer approving everything automates the decision and adds somebody to carry the blame.
How do you tell if oversight is real?
The override rate. If nobody ever overturns a recommendation, the review step exists on paper rather than in the decision.
Where does bias come from?
The data. A system trained on past decisions reproduces past decisions, including the ones you would struggle to defend.
Can good intentions prevent it?
No. Bias is a property of data rather than of design intent, which is why it has to be measured rather than reasoned about.
How do you check for unfair outcomes?
By comparing results across groups in your own decision history, which most organisations have never actually done.
What if the check finds something?
You will have to decide what to do about it. We report what we find rather than what is comfortable, and that is the point of checking.
What should an affected person be told?
The main reason and what would change it. Describing how the model works is accurate and useless to somebody who was refused.
Do people need a way to appeal?
Yes, and it is the part most systems omit. A decision nobody can challenge is a decision nobody can correct.
Does this apply to internal decisions too?
Yes. Recruitment and performance systems decide about people just as lending systems do, and are scrutinised less.
Do you write the principles for us?
No. Plenty of people will, and it will change nothing, which is why your existing statement has not changed anything.
Is this required by law?
Parts of it connect to obligations you already have. The rest is about being able to defend a decision when somebody asks you to.
Would you tell us not to build something?
Yes. Some decisions should not be made this way, and occasionally that is the honest answer to the question asked.
How much decision history do you need?
Enough to see patterns across groups. Where records are thin we will say what the testing can and cannot establish.
Can a bought system be made responsible?
Partly. You control the oversight, the explanations and the appeal route even where you do not control the model itself.
How soon would we see findings?
Three to six weeks, depending mostly on how much decision history exists to test against.
Should customers know they are dealing with AI?
Deciding deliberately is the point. Being found out later is considerably worse than having disclosed it plainly.
Who sees our decision history?
Only the people doing the testing, under whatever confidentiality terms you set, and we work from records rather than needing identities.
What does outcome testing cost?
The number of decision systems and the depth of the testing determine it, quoted after the review call.
Glossary of terms
Responsible AI terms in plain language.
- Testable commitment
- A rule specific enough that a state of the world could break it, unlike a values statement.
- Human in the loop
- A person in the decision path, which only counts as oversight if they have authority and sometimes disagree.
- Override rate
- How often a reviewer disagrees with the system, and the clearest indicator of whether oversight is real.
- Outcome testing
- Checking results across groups to find unfairness that intention cannot rule out.
- Explainability
- What an affected person can be told, judged by whether they can act on it.
- Appeal route
- How somebody challenges a decision, which most systems leave out entirely.
Related services in AI Compliance
Talk to us about responsible AI
Take your published AI principles and ask what would have to happen for each one to be broken. Any that survive that question are not commitments.
Start your project