How to Write an AI Research Paper in High School?

Most high schoolers meet the research paper for the first time as a vague, intimidating assignment: "pick a topic, find sources, write ten pages." Nobody actually teaches you the process, so you improvise one under deadline pressure, and it shows. AI research adds a layer most general writing guides never cover: the paper has to hold up technically, since your model, dataset, and evaluation method all have to be defensible, not just your prose. A research paper is fundamentally different from an essay, because you are investigating a question you do not yet know the answer to, and showing your reasoning, including your model's limitations, along the way.

If you want structured support while you do this, Veritas AI pairs you with a mentor who has written and guided AI research papers before and can catch the weak points in your methodology long before a teacher or judge does.

What Makes a Strong AI Research Paper?

A paper that reads as genuinely researched, rather than a coding project with a few paragraphs of prose stapled on, comes down to three things: a narrow question that names a model, dataset, or technique, a thesis that takes a real, checkable position, and evidence from your own experiments rather than only citing what other papers found. The seven steps below walk through how to get each one right.

Step 1: Narrow Your Topic Into a Specific, Testable Question

Start by picking a subfield you are genuinely curious about, such as NLP, computer vision, reinforcement learning, or interpretability. Then spend time skimming papers on arXiv or Papers With Code until you notice a specific gap or open question. "AI is transforming healthcare" doesn't name anything you could actually test. "Does a sentiment classifier fine-tuned on adult-written text perform worse on teen social media language, measured by F1 score?" gives you something you can actually build and evaluate.

A good test is whether your question names a model, a dataset, and a metric. If your topic is still just a subfield with a question mark attached, it is not narrow enough yet, and narrowing it is the single highest-leverage step in the entire process.

Step 2: Find and Evaluate Your Sources

Once your question is narrow, your search gets much easier, because you are not searching "machine learning"; you are searching something specific like "sentiment classification domain adaptation." arXiv is the best starting point for the actual papers behind almost any AI subfield, and Papers With Code pairs those papers with working implementations, which helps you see what a comparable project actually involves.

Not all sources deserve equal weight in your paper. A peer-reviewed or widely cited paper should carry more weight than a preprint nobody has built on, which should carry more weight than a blog post summarizing either one. If you cannot say who wrote a source, what they tested, or whether others have replicated it, be cautious about building an argument on top of it.

Step 3: Build Your Thesis and Outline

Your thesis is the specific, checkable claim your paper exists to support, stated in a way that names your model, comparison, or measured result. "Sentiment analysis has limitations" is not a thesis. "A sentiment classifier fine-tuned on adult-written product reviews shows a measurable drop in F1 score on teen social media language, even after controlling for vocabulary overlap" is one, because your experiments either support that claim or do not.

Once you have a thesis, outline before you draft. A typical AI paper follows this structure: introduction, related work, a methods section detailed enough that someone else could reproduce it, results, and a discussion of what it means and where your approach fell short. Writing the outline first keeps you from discovering halfway through a draft that your methodology does not actually support your claim.

Step 4: Draft the Body First

Write your methods and results sections before you write the introduction. It is much easier to introduce an experiment you have already run than one you are still figuring out as you write it. Draft in whatever order keeps momentum, then reorder to match your outline.

Momentum matters more than order at this stage. If you get stuck describing your results, move to your methods section, which is usually more mechanical to write, and come back once you have had more time to think about what your results actually mean.

Step 5: Cite and Format as You Go

Most high school AI papers use APA or IEEE format, and citing preprints, GitHub repositories, and dataset documentation correctly is easy to get wrong.

Purdue OWL is the most reliable free reference for standard citation mechanics, and most citation managers, including Zotero, can pull metadata directly from an arXiv link, which saves you from formatting preprint citations by hand.

As you cite sources and describe your methodology, keep a running record of exact model versions, hyperparameters, and dataset versions you used, not just the papers you read. Reconstructing your own experimental setup after the fact is one of the most common ways students lose credibility on an otherwise strong paper, since a reviewer's first real check is often whether your results could be reproduced from what you wrote.

Step 6: Revise With Fresh Eyes

Revision is where a decent paper becomes a strong one, and it is the step students skip most often because the paper already has an ending. Read your draft asking one question: does every section actually support your thesis, or does it include a tangent, like an unrelated model comparison, that does not belong? Cut the tangent, even if you liked writing it.

Get someone with real ML experience to read it, ideally someone who will push back on your methodology, not just your prose. They will catch issues like data leakage, an unfair baseline comparison, or an overstated claim that you cannot see anymore, simply because you have looked at your own experiment too many times to notice where it stops holding up.

Step 7: Decide What Happens After Your First Draft

Once your paper is finished, AI-relevant research competitions give you a place to submit it, and checking their judging criteria before you start writing tells you what a strong submission looks like.

Publishing in an actual journal is often a better path than a competition, since your paper gets a permanent, citable home rather than a one-time result. The Journal of Emerging Investigators, National High School Journal of Science, and International Journal of High School Research all accept AI and STEM research from high schoolers, with different selectivity and fees worth checking before you submit. Check each journal's guidelines before you finalize your draft, not after, since word count and format requirements vary enough to cost you a full round of reformatting.

If you want help matching your paper to the right journal built into the process from the start, that is specifically what the Veritas AIAI Fellowship is structured around: twelve to fifteen weeks working one-on-one with a mentor, often a PhD or PhD candidate in machine learning from a school like Harvard, Stanford, or Oxford, to develop an original applied AI research paper from question to final draft.

Resources

Where to find AI-specific sources

• arXiv: Free access to the papers behind every AI subfield.

• Papers With Code: Pairs papers with working code implementations.

• Google Scholar: Catches older or established work arXiv's recent listings might miss.

Citation and formatting tools

• Purdue OWL: The most reliable free reference for citation mechanics.

• Zotero: Free citation software that can pull metadata directly from an arXiv link.

• EasyBib: A simpler citation generator if you want something faster than a full manager.

Where to publish

• Journal of Emerging Investigators: The most established outlet for hypothesis-driven STEM research by middle and high schoolers.

• National High School Journal of Science: Free to submit and a reasonable first target.

• International Journal of High School Research: Accepts STEM and data-driven research on a rolling basis.

Frequently Asked Questions About How to Write an AI Research Paper in High School

1. How long should a high school AI research paper be?

Most run between 1,500 and 3,000 words, depending on your assignment or venue. Length is not the goal: a tightly argued 1,800-word paper with clear methodology and honest results reads better than a padded 3,000-word paper repeating the same point.

2. What is the difference between an AI research paper and a coding project write-up?

A coding project write-up describes what you built. An AI research paper investigates a specific question you did not know the answer to and uses your model as evidence toward answering it. Your model is a means of testing a hypothesis, not the end goal in itself.

3. How many sources do I need for a high school AI research paper?

There is no fixed number, but eight to fifteen credible sources is a reasonable range. What matters more than the count is the mix: foundational papers that established the technique you are using, and recent papers close to your specific question, rather than a long list that all says the same thing.

4. Can I use a pretrained model instead of training one from scratch?

Yes, and for most high school timelines it is the better choice. Fine-tuning or evaluating a pretrained model from a source like Hugging Face is a legitimate research approach, as long as your methods section is transparent about what you built versus what you started from.

5. How do I pick a topic if I don't know which AI subfield I'm interested in yet?

Skim papers or working code examples across two or three subfields before committing, and notice which one makes you want to keep reading. That reaction is a more reliable signal than reasoning your way into an interest you have not actually tested yet.