14 Biology Project Ideas for High School Students
14 biology project ideas for high school students, with the question, materials, level, time, data and approval rules for each, for summer or the school year.
13 AI project ideas for high school students and beginners, from no-code demos to research papers, with free tools, time needed and how to start.
A good first AI project trains a model on data you collect or download, then tests it on examples it hasn't seen. It ends with something you can show: a working demo, an app or a short paper. Below are 13 ideas grouped by level and field.
For projects outside AI, see our passion project ideas for high school students.
Pick a problem you can get data for, build a first model with a no-code tool, then rebuild it in Python. Test it on data the model never saw and compare the result with a simple baseline.
Good beginner projects use a no-code tool or a short Python notebook and take one to three weeks. The first two below need no code, and the other two need basic Python.
Photograph at least 50 items of each type, such as recycling, trash and compost, and train an image model in Teachable Machine. Then test it on 30 items it hasn't seen. Your photos stay on your computer unless you save the project to Google Drive.
Train Teachable Machine's sound model to recognize a doorbell, a running faucet and your dog's bark. Then test it with sounds it hasn't heard, including similar ones like a phone ringtone. Don't use it in place of a smoke alarm.
Collect 30 to 50 questions students actually ask, like bell times or club sign-ups, and write an official answer for each. Your bot matches each new question to the closest one on your list and returns that answer. If even the best match is weak, have it reply "I don't know; please ask the front office."
To build a bot that writes its own replies with a large language model, read the terms first. Google's Gemini API, for example, requires users to be 18 or older and doesn't allow apps directed at, or likely used by, people under 18.
To launch a full AI app with guidance, our Nova Product Launch program takes you from a problem to a published app, website or game.
Use the MovieLens small dataset, with 100,000 ratings of 9,000 movies by 600 users, to recommend films based on what similar users liked. Add 20 ratings of your own and see if its picks make sense. GroupLens says this version changes over time and isn't meant for reporting research results, so treat it as practice.
A working model is a demo that takes new input during your presentation and answers on the spot. The easiest is the recycling sorter or sound detector above, running on a laptop. Bring a printed chart of its accuracy and three examples it gets wrong.
Social impact projects start with a problem people near you face and measure it with public data or your own recordings. Each ends with numbers you can share, such as forecast errors for your county or error rates by accent.
Download daily air quality data for your county from the EPA's AirData files and train a model to predict tomorrow's AQI from the past week. Compare it with the simplest guess: tomorrow will match today. The files are updated twice a year, so your forecasts will cover past days.
Train an image classifier on PlantVillage, a set of 54,303 healthy and diseased leaf photos in 38 categories. Start from a pretrained image model, following TensorFlow's transfer learning tutorial. Then test it on leaves you photograph yourself and measure how much accuracy drops.
OpenAI's Whisper is an open-source speech recognition model, and its model card notes "disparate performance on different accents and dialects." Record volunteers with different accents reading the same passage, with their permission, and count the words Whisper gets wrong.
An AI research project asks a question about a model or a dataset and answers it with an experiment. You don't need to invent a new model; testing an existing one carefully counts. The three below can become a science fair paper, and our computer science research topics list has more.
The model card for Google's BERT warns that it "can have biased predictions": its top guess for "The woman worked as a [MASK]" is "nurse." Run sentences like "The nurse said that [MASK] was tired" for 50 jobs and record whether "he" or "she" ranks higher. Then compare the results with Bureau of Labor Statistics data on who holds each job.
MedMNIST is a set of 12 two-dimensional and 6 three-dimensional biomedical image datasets, with base images of 28 by 28 pixels. Pick one dataset, train a classifier and study the cases it gets wrong. The site says the data is "NOT intended for clinical use," so your paper should say the same.
Each trading day for a month, log headlines about five companies from one news site in a spreadsheet, then score each with FinBERT. The model was fine-tuned on the Financial PhraseBank, 4,840 news sentences labeled positive, negative or neutral. Then test whether a day's sentiment predicts the next day's price change.
Creative AI projects use a model to make images, music or writing. Keep a log of what it produced and what you changed, so you can explain your part. Use only images, music and text you have the right to use.
TensorFlow's neural style transfer tutorial repaints one image in the style of another. Use your own photos as content and public-domain paintings from the Met's Open Access collection as styles. Make a series and write about which styles worked and why.
Magenta RealTime 2 is an open-weights music model you play like an instrument. Record a piece where you and the model trade phrases.
TensorFlow's text generation tutorial trains a character-level model on Shakespeare's writing. Generate a short scene and stage a reading with friends. Then retrain the model on a public-domain author you like and compare the two styles.
The most common mistake is testing a model on its own training data, which makes its accuracy look far better than it is. Plan for these other problems in week one, too:
Science fairs, app competitions and a federal K-12 challenge all take AI projects. Two of the three below start at the district or regional level, so begin there.
Facts checked October 2026. Dates come from each program's official page.
Yes: the Teachable Machine projects need no coding, and in the US you can manage your own Google account from age 13. Younger students can use an account a parent sets up through Family Link. Some services start at 18, including Colab's built-in AI features. For ideas sized for younger students, see our passion project ideas for middle school students.
Only for the music idea (12), which needs a Mac with Apple Silicon. Google Colab runs the work on Google's machines. Its free GPUs are limited and not guaranteed, so start with small datasets like MedMNIST or MovieLens small.
Not much to start. The beginner projects above need only percentages to read your results. Statistics helps when you compare models, and linear algebra and calculus matter once you build neural networks from scratch.