There is a curious assumption developing around artificial intelligence: if an organisation can use AI for something, it probably should.

We do not agree.

AI is becoming an important part of software development, healthcare and education. It can solve problems that were difficult or impractical only a few years ago. But that does not make it the right solution to every problem.

At Virtual Medical Coaching, our approach is deliberately selective.

The question we start with is not, “Where can we add AI?”

It is, “What are we trying to achieve, and what is the most reliable way to achieve it?”

Why does VMC use AI selectively?

Healthcare education is not a good environment for technology that has been added simply because it is fashionable.

Learners need systems that behave consistently. Educators need to understand what is being assessed. Scientific relationships need to remain scientifically correct. Where a result can be calculated from known information, we generally do not need a generative AI model to guess the answer.

Consider radiography and radiation safety.

Many consequences of a learner's actions are governed by established relationships. Exposure parameters, distance, positioning, collimation, geometry and radiation protection decisions have effects that can be calculated or modelled.

Those relationships do not become better because a large language model has been inserted between the learner and the result.

In situations like these, deterministic software can be more appropriate because the same action should produce the same consequence.

Responsible use of AI can sometimes mean deciding not to use AI.

When does VMC use AI?

AI becomes valuable when the problem is not adequately solved by a fixed rule or traditional software alone.

This can include areas such as identifying patterns in learning, adapting educational material, interpreting more complex learner interactions, assisting with the creation of synthetic information, or helping determine what learning experience could be useful next.

The distinction is important.

AI is not the educational strategy. It is one technology available within the educational system.

Our adaptive learning work is a good example. A learner may demonstrate strength in one area while repeatedly struggling with another. Software can use that information to help determine what they should practise or review next.

Some of that process can be deterministic. Some can use statistical models. Some may benefit from AI.

We do not need to hand the entire learning pathway to a generative model simply because generative AI exists.

Why does VMC run AI on its own GPUs?

Another important question is where an AI model actually runs.

When someone interacts with an AI system, information may be sent to infrastructure operated by another company. That arrangement may be completely appropriate for some applications, but it should not be an automatic architectural decision, particularly when software is being developed for healthcare and education.

VMC has invested in the ability to run AI models on our own GPU infrastructure.

This gives us more control over the model, the information supplied to it, and the role it is allowed to perform.

Running models on infrastructure we control can also reduce unnecessary dependence on external AI services. It allows us to decide when information needs to leave our environment and, just as importantly, when it does not.

This does not mean that every AI workload must always run locally. Different technical problems have different requirements.

It means that the decision remains ours.

Does VMC allow AI to make clinical decisions?

VMC develops educational simulation software. Our AI systems are not intended to replace clinicians, educators or professional judgement.

A simulation may need to evaluate what a learner has done, determine the consequence of an action, select suitable educational material or provide feedback. Those are educational functions.

Where clinical or scientific accuracy matters, AI operates within a wider system of defined rules, subject matter expertise, testing and human review.

A response that merely sounds convincing is not sufficient.

This is particularly important with generative AI. Large language models are extremely good at producing fluent language. Fluency and correctness are not the same thing.

Our job is therefore not simply to connect a learner to an AI model. It is to determine what the model is allowed to do, what information it can use, how its output is constrained and where another system or a person needs to remain in control.

How does AI fit into medical simulation?

Simulation provides an interesting environment for AI because the objective is not merely to provide an answer. The objective is to help someone learn.

Sometimes the best learning experience is highly structured.

A learner performs a procedure, makes a decision and sees the consequence.

Sometimes the useful next step depends on what that individual learner has already demonstrated.

That is where adaptive systems can become valuable.

AI may help us create more personalised educational experiences, but personalisation still needs an educational purpose. Making software unpredictable is not the same as making it adaptive.

How is VMC using AI in CT simulation?

Computed tomography is one area where AI techniques can create possibilities that conventional development approaches make difficult.

For example, AI can contribute to the creation of synthetic imaging information that can then be incorporated into an educational simulation.

That is a fundamentally different use of AI from asking a chatbot to explain CT.

The AI is helping solve a specific technical problem inside a larger simulation system. The surrounding educational experience can still use conventional software, defined physics, structured tasks and educator oversight.

This combination is much closer to how we see AI developing within VMC.

What does human oversight of AI mean at VMC?

Human oversight is sometimes discussed as though a person simply needs to inspect an AI response before it reaches a user.

We think it starts much earlier.

People decide whether AI should be used at all.

People define the educational objective.

People determine what information the model receives.

People decide what the model is permitted to influence.

Subject matter experts review the clinical and scientific content surrounding the system.

And where deterministic software is more appropriate, people make the decision to use that instead.

Is this what responsible AI in healthcare education looks like?

There will not be a single architecture that defines responsible AI.

The World Health Organization has emphasised that AI for health should be designed and used with appropriate governance, accountability and protection of human interests. The US National Institute of Standards and Technology has similarly developed its AI Risk Management Framework to help organisations identify and manage risks associated with AI systems.

Those principles matter, but responsible deployment also involves ordinary engineering decisions.

Does this feature need AI?

What happens when the model is wrong?

Can a known rule provide the answer instead?

Where does the data go?

Who controls the infrastructure?

Can someone explain why the system behaved the way it did?

These questions are less exciting than announcing that every product is now “AI-powered”.

They are also much more useful.

Will VMC use more AI in the future?

Almost certainly.

AI capabilities are developing rapidly, and some of those developments will allow us to build educational experiences that were previously impractical.

But increased capability does not change the principle.

We will use AI when it makes the simulation, the learning process or the underlying technology better.

We will use conventional software when conventional software is the better tool.

We will keep control over where AI sits within our systems.

And where appropriate, we will keep the models themselves running on infrastructure we control.

AI should be a tool inside the learning system, not the reason the learning system exists.

That may mean VMC uses less AI than some companies.

We are comfortable with that.