RAINThe Oath Meets the Machine, Ethics Speech
The Oath Meets the Machine
Oct 15, 2026 ·
Opening
Slide 1.
@olusola sayeed ayoola
For the last thirty minutes, we talked about what AI gets wrong. For the next thirty, I want
to talk about something harder: what we owe the people whose lives feed it.
Slide 2.
Let me begin with good news. Every one of you in this room has already taken an oath.
Confidentiality. Consent. First, do no harm.
So the ethics of AI is not a strange new subject for medicine. You already know the duty. AI
does not change the duty. It simply makes it much easier to break, and much harder to
notice that you have broken it.
Slide 3.
Do you remember the patient from the last session? The one who said, I'm fine, doctor,
while his search history told the truth?
Now imagine his file. His name, his history, his results. Pasted into a free public chatbot by
a busy doctor at the end of a long clinic, who only wanted help writing a referral letter.
In that instant, a set of questions opens that nobody in the consulting room can answer.
Whose server is this data now sitting on? In which country? Who will see it? And for how
long will it be kept?
This talk is about those questions. And it ends with a decision only we can make: the rules
this College will live by.
Privacy
Slide 4.
The first instinct of a careful doctor is to remove the name. Good. But removing the name
is not enough.
You may delete the name and forget the National Identification Number. You may remove
the NIN and leave the health insurance number. You may remove both and leave the
hospital number, the phone number, the face in the CT scan. Or the voice on the
recording. Any one of these can lead straight back to the person.
Page 1 of 7
The Oath Meets the Machine, Ethics Speech
This is why anonymisation is not a clerical task. Only the expert who knows the data
knows what still points to the patient, and therefore what must be removed.
Slide 5.
Let me show you how little it takes. In the year 2000, Professor Latanya Sweeney at
Carnegie Mellon University showed that 87 percent of Americans could be uniquely
identified by just three details. Their postcode. Their date of birth. And their sex. No name.
No ID number. Three ordinary facts.
Now think of our own equivalents. The ward. The age. The hometown. A rare diagnosis. Put
two or three together in a small community, and everybody knows who you are talking
about.
Slide 6.
So what is the solution? Code people. Don't name them.
Inside the hospital, Mrs Bello remains Mrs Bello. But the moment her record goes
anywhere near an AI, she becomes P-0417. The key that connects P-0417 to Mrs Bello
stays locked inside the institution. The AI can study the case as deeply as it likes. It simply
never knows whose case it is.
This is not a technology problem first. It is a records problem. It starts with how we
digitise, how we keep records, and how we refer to the people we serve. An institution
that still runs on paper folders and full names cannot do this. An institution that has
digitised and coded its records can.
Slide 7.
Now, your voice is a fingerprint.
There are AI tools today that sit quietly in the consulting room, listen to the conversation
between doctor and patient, and write the notes. They are genuinely useful. But voice is
biometric data. Remove the name, and the voice can still identify the speaker.
So we must ask: where is that recording kept? For how long? Who can hear it?
And there is a deeper, more human question. If a patient knows she is being recorded, will
she still tell you the real complaint? Will the young man still confess the thing he came to
confess? Even if we promise not to keep the voice, can she be confident enough to speak
freely? That is not a technical question. That is the doctor-patient relationship itself.
Slide 8.
Here is something many people do not realise. Data used to train AI does not go straight
into a cold machine. It passes through human hands.
Page 2 of 7
The Oath Meets the Machine, Ethics Speech
Around the world, millions of people, many of them young, work as data annotators. They
label pictures, sort text, check the AI's answers, rate its responses. They are paid per task.
They work in Kenya, in the Philippines, in India, in Venezuela, and yes, right here in Nigeria.
They are usually not doctors. They are simply told what to look for.
Most work under confidentiality agreements, and most have no interest in your patient at
all. But what if one does? What if your patient is a public figure, and a stranger somewhere
reads her diagnosis while labelling data? A machine does not gossip. A person might.
Slide 9.
So let me answer the question some of you are quietly asking. If I have already pasted
patient information into ChatGPT, or Gemini, or Claude, or Meta AI, can I take it back?
Partly. You can delete the conversation, and the provider will remove it from its systems
over the following weeks. OpenAI, for example, says deleted chats are permanently
deleted within about thirty days, unless the law requires them to keep it longer. You can
switch off model training in the settings, so that future conversations are not used to train
the model. And on paid business accounts and through the developer interface, data is
generally not used for training by default.
But here is the hard truth. If your data has already been used in a training run, it cannot be
taken back out. There is a research field called machine unlearning, which tries to make
models forget. It is research, not a guarantee. There is no button that makes the machine
un-know what it has learnt.
Slide 10.
So if you have done it, and many good doctors have, here is what to do. No shame, only
action.
Delete those chats today. Switch off model training. Move your work to an institutional
account. And never again paste anything that identifies a patient.
That last step is the only real protection. The safest data is the data you never sent.
Consent and responsibility
Slide 11.
Now let us talk about consent.
When a patient lies on your couch, opens her mouth, rolls up her sleeve, she is giving
consent. But to whom? To a human being. To you. She did not consent to an algorithm.
Page 3 of 7
The Oath Meets the Machine, Ethics Speech
And that tells us something important. If consent was given to a human, then a human
must carry full responsibility for whatever the AI contributed. AI can only ever be the
assistant. Never the doctor.
The human signs. The human is responsible. You can never stand before a patient, a
journal editor or a court of law and say, the AI said so.
Slide 12.
It follows that clinical AI tools are for the practitioner, not the patient.
Of course, patients will use AI. We cannot stop them, and we should not try. But we must
teach them that what a chatbot tells them is private research, never authority. They must
always bring it back to their doctor.
Some applications now attach a real doctor who reviews every AI answer, so the patient
knows a human has seen it. That helps. But there is a deeper problem. A symptom is never
just words. When a patient types, this is what I am going through, that is only the
beginning. Diagnosis needs examination. It needs the laboratory. It needs the test.
Slide 13.
And if AI is now doing the test itself, we must be even more careful.
There are algorithms that photograph a blood smear and look for malaria parasites. There
are claims of reading blood pressure from a single pattern. These may be wonderful. But
we must prove them. Were they validated on our slides, our parasites, our microscopes,
our patients? What conditions make them fail? And if they fail, would anybody notice?
That last question is the dangerous one. A test that fails silently, with no human in the
loop, does not fail once. It fails on every patient, the same way, until someone finally looks.
Slide 14.
So every medical AI must be vetted before it is adopted. The scribe that listens to the
consultation. The adviser that suggests a diagnosis or a treatment. The reader that
interprets the ECG, the chest X-ray, the mammogram.
Before any of these enters our hospital, we ask four questions. Was it tested on our
population? Where does the data go? Who is accountable when it is wrong? And can we
audit it? If a vendor cannot answer all four, the tool does not come in.
Sovereignty and capacity
Slide 15.
Let me ask a simple question. Where does your data sleep at night?
Page 4 of 7
The Oath Meets the Machine, Ethics Speech
When we compute with AI, the data must live somewhere: a database, a cloud, a server. If
that server sits in another country, under another country's laws, then another country's
rules decide what happens to our patients' information.
So our data should live in sovereign storage, where our laws and our institutional rules
apply, and where we can dictate how it is managed. It should be encrypted everywhere:
when it is stored, and when it travels. Encryption means that even if data ends up in the
wrong place, in the wrong hands, it is useless to them.
And we do not have to invent the law. It is already here. The Nigeria Data Protection Act of
2023 classifies health data and biometric data as sensitive personal data, and the Nigeria
Data Protection Commission enforces it. Our College policy should stand on that
foundation.
Slide 16.
Now, I know the objection. Compute is expensive, and we do not have enough
infrastructure. That is true. So we must be clever.
First, edge computing. With TinyML and related techniques, we can shrink AI models until
they run on small, cheap, low-power devices, even without internet. The data never has to
leave the room.
Second, transfer learning. We do not need to train giant models from scratch. We can take
a model that has already been trained and teach it our own local data, on modest
machines that we own.
Third, offline models. Open-weight models, such as Meta's Llama family, can be
downloaded and run entirely on our own servers. Nothing goes out. And the large cloud
datasets still have a role: to validate and test what we have built and learnt locally.
Slide 17.
Which brings me to a conviction I hold deeply. Off-the-shelf rents intelligence. Building
owns it.
Off-the-shelf tools are easy, cheap and accessible. We will use them, and we should. But
an institution like this one cannot live on off-the-shelf alone.
We must invest in bespoke solutions. We must employ machine learning developers of our
own. People who build our data collection pipelines between hospitals and patients.
People who train models on our data, whose algorithms we can see, monitor and protect.
People who watch our cloud subscriptions and deployments to make sure they comply
with our laws. And people who uphold ethical standards that we decide, here, locally.
Build internal capacity. Train locally. Own it. That is how we will win this race.
Page 5 of 7
The Oath Meets the Machine, Ethics Speech
Institutional governance and closing
Slide 18.
So how does an institution govern all of this? Individual caution is not enough. We need a
policy. And every strong policy begins with red lines: rules with no exceptions, that
nobody has to weigh in the heat of a busy clinic.
I propose two.
Red line one. No identifiable patient data goes into a public AI tool. Not once. Not because
it was urgent. Ever. The moment we allow "just this once", the line becomes a slope.
Red line two. The human signs, and the human is responsible. AI advises. It never
prescribes.
Slide 19.
Around those two lines, four working rules, which can change as the technology changes.
Approved tools. The College should not leave staff to use whatever AI they can afford. It
should provide institutional subscriptions, under clear rules, so staff have good tools and
the institution keeps control.
Disclosure. When AI was used in research or teaching, we say so. AI must never be a silent
ghostwriter.
Competence first. Nobody uses AI on real work until they have been trained on its limits,
which is exactly why we are here today.
A living policy. A small standing committee reviews the rules every six months, hears
problems, and updates the list of approved tools.
Slide 20.
And here is what I most want this College to do. Do not ban AI. Study it.
Create space for staff to come together and discuss how AI actually performed. And
record everything. What we thought. What the AI thought. What actually happened. Over
time, that record becomes something no vendor can sell us: our own evidence of where AI
helps and where it does not. Our own understanding of its applicability, its
appropriateness and its suitability, in our setting, for our patients.
That, again, depends on digitalisation. You cannot learn from what you did not record.
Slide 21.
There is one more ethical question we cannot avoid. Who owns it?
Page 6 of 7
The Oath Meets the Machine, Ethics Speech
Who owns the patient's data? Who owns the model trained on that data? Who owns the
knowledge it produces? These questions must be settled before the project begins, in the
consent form and in the contract. Not after the dispute.
Slide 22.
But let me be careful not to leave you with fear alone. Privacy is not the enemy of science.
We need AI. We need to study the human being: our physiology, our anatomy, our
psychology. We need to understand how every treatment, every habit and every event in
a life relates to the illness that eventually arrives.
So while we protect everybody's data, we should also honour and encourage the people
who are willing, with full and informed consent, to become open specimens for life-long
study. People who allow their whole medical life to be linked and learnt from. We must
build that community, protected as far as possible, because that is how science grows.
Slide 23.
In my last session, I left you with three words. Relate. Rely. Respond.
Let me add a fourth. Protect.
Protect the patient, whose trust we hold. Protect the student, whose mind we are
shaping. And protect the name of this great College, which has trained some of the finest
physicians this country has ever produced.
Slide 24.
In a few minutes, we will sit together to build our institutional AI guidelines. I have offered
you two red lines and four working rules. The decision is yours.
Let us write them well. Thank you.