Resource Guide
Research Project Ideas in AI & Machine Learning for High School Students
Specific, answerable questions across models and methods, ethics and society, applied machine learning, and the foundations of intelligence, with guidance on choosing one.
How to Use This List
Artificial intelligence is the most over-subscribed and over-hyped field a student can choose, which is precisely why a careful, honest project stands out. Admissions readers and interviewers have seen countless applicants claim to have "built an AI". Far rarer, and far more impressive, is a student who can explain exactly what their model did, why it worked or failed, and what that actually demonstrates.
The questions below are framed to produce that kind of understanding. Take one that genuinely interests you and narrow it until it is specific and testable. Our guide to writing a research question is the place to start.
Choose a Method Before a Topic
The method is the decision that determines whether a project can actually be finished; the topic is secondary. Every question below is tagged with one of these four routes, its difficulty, and the finished output it should produce.
Build and benchmark
Implement something and measure it. AI is a subject where a laptop produces real evidence, and the difference between a strong project and a weak one is measurement: state your inputs, your timing method and your variance, not just that it ran.
A language you know and a question that has a number for an answer.
Open data analysis
Train or evaluate on published data rather than data you collected. Air quality, transport and epidemiological datasets are all open and largely unused by students.
One dataset, a notebook, and an honest baseline to compare against.
Literature review
For questions you cannot benchmark, read the primary papers and lay out what is established and what is contested. Most of the AI literature is free on arXiv, which is unusual and worth exploiting.
arXiv, and the discipline to read the method section rather than the abstract.
Ethics argument
Defend a position on a real deployment rather than on AI in general. The abstract version of this question has been written thousands of times; the specific version has not.
One deployed system you can describe precisely, and the strongest argument against your view.
Ideas by Sub-Field
Models & methods
-
How does training-data size affect a model’s performance on a task, and where do the gains plateau?
Build and benchmark Advanced
Output: A scaling curve you produced, with the plateau identified and run-to-run variance reported
-
Can a small model trained on open data match a larger one for one narrow, well-defined task?
Build and benchmark Advanced
Output: A head-to-head benchmark on one narrow task, with compute cost stated alongside accuracy
-
How do different ways of handling imbalanced data change a classifier’s accuracy and fairness?
Build and benchmark Intermediate
Output: A comparison of resampling methods on one dataset, reporting both overall accuracy and per-group error
AI, ethics & society
-
How does bias enter a machine-learning system, and which mitigation methods actually reduce it?
Literature review Intermediate
Output: A review of documented bias cases and which mitigations measurably reduced the gap
-
How should a medical or legal AI’s accuracy be weighed against its interpretability?
Ethics argument Advanced
Output: A position paper on one deployed system, stating what it should be required to explain and to whom
-
What does the research say about the environmental and energy cost of training large models?
Literature review Intermediate
Output: A synthesis of published training-cost estimates and a clear account of what they leave out
Applied machine learning
-
Can a model meaningfully predict a public outcome, such as air quality or disease spread, from open data?
Open data analysis Intermediate
Output: A model built on open air-quality or epidemiological data, compared against an honest naive baseline
-
How well do sentiment-analysis methods actually capture meaning in informal or sarcastic text?
Build and benchmark Intermediate
Output: An error analysis of the cases your classifier gets wrong, grouped by failure type rather than listed
-
How robust is an image classifier to small, deliberate changes in its input?
Build and benchmark Advanced
Output: A robustness test using perturbations you generated, with the accuracy drop quantified
Foundations & reasoning
-
What does it mean to say a language model “understands”, and what do current systems actually do?
Literature review Advanced
Output: An essay setting benchmark claims against documented failures and arguing a defended position
-
How do reinforcement-learning agents learn, and in what situations do they reliably fail?
Literature review Advanced
Output: A review of documented failure modes organised by underlying cause rather than by paper
-
How can we tell whether a model has genuinely generalised rather than memorised its training data?
Build and benchmark Advanced
Output: A held-out test you designed specifically to separate generalisation from memorisation, with results
Doing AI Research Honestly
The fastest way to weaken an AI project is to overclaim. A model that reaches 92 per cent accuracy on a tidy dataset has not "solved" anything; the interesting questions are what the remaining errors look like, whether the result holds on messier data, and what the model is really keying on. A project that investigates those questions honestly is far stronger than one that reports a headline number.
Good practice also means being careful about data: where it came from, whether it is biased, and whether using it raises privacy concerns. These are not obstacles to a good project; they are often the most interesting part of it.
For projects that are more about classical computing than learning systems, see our companion guides to research project ideas in computer science and engineering.
Free PDF
Take the 50 strongest ideas with you
A free PDF with fifty of these questions tagged by method and difficulty, the output each should produce, and the feasibility checklist we apply before recommending any project. We email it to you, along with occasional guidance worth having. No obligation, and your details are not passed to anyone.
Taking a Question Further
AI reaches into almost every field, and the strongest projects often pair it with a real domain: medicine, economics, law, the environment. For the wider context, see our Technology, AI & Engineering field page and our broader research project ideas across all six fields. When you are ready to turn a question into a finished project with a mentor who works in the field, the Research Scholar programme is built for exactly that.
Frequently Asked Questions
Do I need to know how to code to do an AI research project?
It helps, but it is not always essential. Strong projects exist on the conceptual and ethical side of AI that involve little or no code. Where a project trains or tests models, a mentor helps a student build exactly the programming they need rather than assuming fluency in advance.
Can I train a model without expensive hardware?
Yes. Free cloud notebooks such as Google Colab and Kaggle provide enough compute for genuine machine-learning experiments on open datasets. Many strong projects use small models deliberately, because the question is about behaviour and trade-offs, not raw scale.
Is it better to build a model or to study AI critically?
Both are legitimate research. A technical project tests a hypothesis about how models behave; a critical project examines bias, interpretability, or societal impact with rigour. The strongest applicants often combine the two: building something, then thinking carefully about what it does and does not show.
How does an AI research project help with university applications?
Computer science and AI courses are heavily oversubscribed and look for genuine understanding beyond hype. A focused project, and the ability to explain honestly what a model did, why, and where it failed, demonstrates exactly the maturity selective courses want.
Now placing students for the October start. Places in each field are limited by mentor availability. We review applications on a rolling basis.
ScholarBridge matches students with doctoral-level or equivalent research mentors across six academic fields. Every project is student-led and completed to a standard the student can stand behind in any university interview.
Explore all programmes