Can HK Secondary Students Actually Build AI? Yes — But Not Like That.

The MediSafe story generated a response I find frustrating, because it conflated two entirely different things: whether that specific project was genuinely student-built (almost certainly not), and whether Hong Kong secondary students can build meaningful AI projects at all (yes, with the right conditions).

Those are separate questions. The scandal's aftermath has muddled them in a way that isn't useful for students, parents, or schools.

Let me try to separate them properly.

What Form 4 students genuinely cannot do alone

I described in an earlier piece what a sophisticated AI medical application actually requires. Let me be more precise about the developmental reality here, because precision matters.

A student who has been learning ICT since P5, who has had a good secondary school computing curriculum, and who spends significant personal time on coding could, by Form 4 or Form 5, be a capable programmer. They might be writing functional web applications, working with APIs, building apps that do genuine things. This is possible. I've taught students like this.

What they cannot do — not because they're not smart enough, but because they haven't had the prerequisite exposure — is train machine learning models on domain-specific medical data with appropriate validation methodology. This requires probability and statistics at a university level, medical domain knowledge, access to appropriate datasets, and the kind of iterative debugging and evaluation process that takes weeks or months even for professionals. There is no fourteen-week school project pathway to this outcome.

The tell, always, is the documentation. Real ML projects built by individuals learning as they go have rough edges. The model selection rationale is tentative. The evaluation metrics are sometimes chosen incorrectly. The data preprocessing has visible seams. A Form 4 student who has actually built an ML system will have learned a great deal and will show work that reflects genuine discovery — including the blind alleys and the things that didn't work.

A ghost-built project looks finished. That's exactly the problem.

What secondary students actually can build

Here's where I want to push back against the defeatism that can come from the MediSafe conversation.

I have seen genuine student AI projects at secondary level. Not enterprise-grade applications. Not research-quality systems. But real, student-owned work that demonstrates understanding, effort, and genuine learning.

A Form 5 student I taught built a classifier that could distinguish handwritten Chinese characters — using a public dataset, a pre-built machine learning framework, and about four months of after-school work. The accuracy was modest. The code was genuinely hers. She could explain every design choice and every failure mode. When it was presented at a school showcase, she spent forty minutes in conversation with judges answering questions I hadn't prepped her for. That's what student work looks like.

Another student built a sentiment analysis tool for Cantonese social media text — a genuinely interesting problem because most off-the-shelf NLP tools are built for Mandarin or English. He used transfer learning from a Cantonese language model, adapted it with a small labelled dataset he collected himself, and documented the limitations honestly. The project won nothing. It was excellent.

The difference between these projects and ghost-built ones is not the polish. It's the ownership. These students could be questioned at any depth about any aspect of their work. That's the test.

What meaningful STEM education requires

The competition trophy system rewards outputs. Real STEM education builds capability — and capability requires a specific set of conditions that not all Hong Kong schools are providing.

Time to fail. The most important learning in any engineering or computational project happens in the failures. The code that doesn't run, the model that predicts nonsense, the approach that seemed right and turned out to be wrong. Students who are given the real experience of debugging under genuine uncertainty develop something that cannot be ghost-built: the ability to work through problems they haven't seen before. This requires projects with enough timeline to have failures and recover from them.

Teachers who know the domain. ICT instruction in Hong Kong is variable in quality. A teacher who can genuinely assess whether a student's machine learning project reflects real understanding — who can ask the right questions during development and catch misconceptions early — is not everywhere. Teacher development in computing at secondary level is one of the genuine gaps in HK's STEM ambition.

Scope matched to level. A genuinely challenging project for a Form 4 student is not a medical AI app. It's a classification system built on a public dataset, a recommendation algorithm for a simple use case, a basic natural language tool. These are real engineering tasks. They produce real learning. They're achievable without professional support. The scope just has to be honest.

Assessment that probes process, not product. The competition system's fixation on polished final products is part of what created the ghost-building incentive. A competition structure that requires working documentation — incremental commits to a version-controlled repository, weekly development logs, a debugging journal — would be much harder to fabricate and would reward genuine process.

What parents should look for

If your child is involved in a STEM programme or submitting to competitions, here are the questions worth asking.

Can your child explain the technical choices in the project? Not just "we used machine learning" but "we chose a decision tree over a neural network because our dataset was small and we needed interpretability." If they can't explain the choices, they didn't make them.

Can your child describe what didn't work? Real engineering projects have failures. A student who can only describe successes either had professional assistance or is being coached to present selectively. Ask: "What broke? What did you try that didn't work? What would you do differently?"

What did your child's teacher see during the development process? A teacher who was engaged with the project in progress should be able to describe the student's development — where they struggled, what they learned, how their understanding changed. If the project appeared fully formed near the submission deadline, something is wrong.

Is the school tracking competition outcomes against actual computing curriculum performance? The student who wins a major AI competition but is getting Cs in secondary computing has a profile worth examining.

The legacy this could have

The MediSafe scandal, if it leads to genuine reflection rather than isolated scapegoating, could produce something useful: a forcing function on schools and competitions to define what "student work" means in an era when professional AI tools and ghost-building services are both available.

That definition is overdue. The education system has been operating as if the meaning of "student-built" is self-evident. It isn't any more, and the gap between what the label claims and what the reality often is has been growing for years.

What I'd like to see: competitions that require live technical demonstration, with judges who can probe. Schools that distinguish clearly between "AI-assisted" and "AI-built" and are honest about which they're seeing. Parents who value genuine capability over credentials, even when credentials are what the university application system is currently counting.

Hong Kong secondary students can build real things with AI. I have watched them do it. The question is whether the system is designed to recognise genuine work — or just to reward whatever shows up polished, regardless of who actually built it.