High-school students still want careers in technology, but they no longer believe there is only one way in. With AI now able to write and refactor large amounts of code — and no AP course in 'vibe coding' — teenagers are actively rethinking which skills actually matter, and their teachers are scrambling to keep up.

From coding to statistics

Educators like Benjamin Rubenstein, assistant principal at Manhattan Village Academy in New York, describe a clear swing away from stacking as many computer-science classes as possible and towards statistics, data literacy and applied maths. If AI can write the code, the differentiator becomes understanding what the numbers mean and being able to argue about them.

The pipeline is now a network

Forty years ago the aspirational path ran through NASA into physics and engineering; twenty years ago Google pulled the same students into computer science. AI is reshaping the picture again, pushing students toward roles that combine computing with analysis, interpretation and judgment — the parts machines still struggle with.

What the numbers show

The shift is visible in the data. Computer-science, computer-engineering and information degrees in the US and Canada fell about 5.5% in 2023–2024, according to the Computing Research Association. Meanwhile roughly 264,000 US students registered for AP Statistics in 2024, making it one of the most popular AP exams — comparable in scale to AP Computer Science Principles and A combined.

AI in the classroom

Educators are also experimenting with AI as a teaching partner rather than a threat. Multi-agent classroom systems being trialled by researchers such as Xiaoming Zhai at the University of Georgia interact with students to model scientific inquiry, flag which students understand a concept and suggest tailored data projects — pointing to a version of high-school STEM where the tools help personalise the learning rather than replace it.

What schools need to do

Curricula have to catch up quickly: less rote coding practice, more model-building, more real datasets, and much more emphasis on interpreting and critiquing AI output. The subject list matters less than the habits of mind it produces.