Steps AI

Don’t just use AI. Train one.

Steps AI is a hands-on course on artificial intelligence and machine learning for grades 5–8. In two sessions, students train an image model, test a voice AI until it breaks, and learn to write prompts that get better answers.

See the two sessions Try the first lab

Is this photo a cat or a dog?

Switch on what the photo shows and watch the network decide.

Cat95%
Dog5%

A toy network with hand-set weights. A real one learns them from thousands of labeled photos.

Grades 5–8
A course for the next generation
2 sessions
1.5 hours each
Free tools
Everything runs in a browser
Hands-on labs
Train, test, prompt and design

Session one builds the foundations. Session two asks the harder questions.

Each session runs 1.5 hours. Every tool is free and runs in a browser.

Session 1

AI & ML Foundations

  • What is AI?
  • AI vs ML vs deep learning
  • Types of learning
  • Neural networks
  • Model training
  • Computer vision
  • Voice recognition

Labs: Sort It!, Teachable Machine, Whisper demo

Session 2

Language, Reasoning & Ethics

  • Natural language processing
  • Transformers and LLMs
  • Prompt engineering
  • Reinforcement learning
  • Modern AI landscape
  • Responsible AI and ethics
  • Capstone project

Labs: Prompt engineering lab, Design challenge, Presentations

Ideas students leave able to explain

Artificial intelligence Machine learning Deep learning

AI, machine learning and deep learning are nested

AI is the broadest idea and has been around since the 1950s. Each ring is a more specific way of getting there.

Artificial intelligence

Any technique that makes computers mimic human behaviour.

Siri, Tesla Autopilot, AlphaFold

Machine learning

Algorithms that improve automatically from data.

Spotify recommendations, Gmail spam filter

Deep learning

Multi-layer neural networks, the most powerful subset.

GPT-4o, DALL·E 3, Gemini, Whisper

Intelligent systems perceive, reason and act

“Doing the right thing when you don’t know exactly what the right thing is.”Peter Norvig, on defining AI

Perceive

Gather information with cameras, microphones, sensors and text.

Visual perception, speech recognition

Reason

Make sense of what was sensed, like recognizing a red light.

Decision-making, planning

Act

Turn a decision into a response, like a self-driving car braking.

Translation, autonomous control

Machines learn in three ways

Supervised learning

Learns from labeled examples and maps input to output.

Spam detection, image classifiers

Unsupervised learning

Gets no labels and finds hidden structure on its own.

Clustering, anomaly detection

Reinforcement learning

An agent acts in an environment and learns to maximize reward.

AlphaGo, ChatGPT RLHF

How a model is trained

  1. Collect and label dataSplit it into a training set and a test set.
  2. Forward passInput goes in and a prediction comes out.
  3. Calculate lossHow wrong was the prediction?
  4. BackpropagationSend the error backwards through the network.
  5. Update weightsTake a gradient descent step, then repeat.
loss training passes
The cycle repeats until the model is as accurate as it can get. In the Teachable Machine lab, students watch the loss fall as it trains.

Labs where students build something, then try to break it

Session 1

Train your first AI model

With Teachable Machine, right in the browser.

  1. Open the siteChoose Image Project, then Standard image model.
  2. Create 3 classesThumbs up, thumbs down and peace sign. Record examples of each with a webcam.
  3. Train the modelClick Train Model and watch the loss go down in real time.
  4. Test and break itTry to fool your model. What confuses it? What works from across the room?

Open Teachable Machine

Session 1

Can AI transcribe you accurately?

A voice AI turns a waveform into a spectrogram, then an encoder and decoder turn that into text. Students test it with OpenAI Whisper.

Challenge A: the Whisper test

  • Open Whisper on Hugging Face.
  • Record yourself reading this sentence:The quick brown fox jumps over the lazy dog
  • Try again with a different accent or speed.
  • Did it get every word right?

Challenge B: fool the AI

  • Record yourself in a noisy place.
  • Try whispering, then shouting.
  • Say a word from another language.
  • Try technical jargon or made-up words.
  • What breaks it? What doesn’t?

Open Whisper on Hugging Face

Session 2

The prompt is your code

Students try these with ChatGPT or Claude and compare the results.

Level 1: a bad prompt

Explain photosynthesis

What’s wrong with this? How could you improve it?

Level 2: a better prompt

Explain photosynthesis to a 6th grader using a food analogy. Keep it under 5 sentences.

The highlights add an audience, a style and a length. How did that context change the output?

Level 3: your turn

Pick any topic. Write the worst prompt you can, then rewrite it as the best. Share the pair with the group.

What students found in class

Data decides accuracy.With only a few samples of each hand sign, the image models gave uneven results. Lighting, camera angle and the amount of data all changed the outcome, and too little data left a model confused.

Voice AI has limits.Whisper handled clear speech well. It struggled with non-English words and with noisy conditions, which is exactly what the challenges were built to show.

How machines learned language

Before 2017, AI read text one word at a time, like a book you can’t flip back through. Then researchers let the model look at every word at once.

  1. 2017

    The transformer architecture arrives with Google’s paper “Attention Is All You Need”.

  2. 2019

    BERT and GPT-2 become the first large pre-trained language models.

  3. 2022

    ChatGPT reaches 1 million users in 5 days, aligned with human feedback.

  4. 2024

    GPT-4o, Gemini 1.5 and Claude 3 handle images, audio and video as well as text.

  5. 2025

    Agents and reasoning models plan and carry out multi-step tasks.

The key idea: self-attention lets a model weigh which words matter most for understanding every other word in context.

How a language model writes

The best thing about AI is its ability to learnThe model turns words into numbers, scores every possible next word and picks a likely one.

The best thing about AI is its ability to learn fromIt feeds the result back in and repeats until the sentence is finished.

Learning from rewards

In reinforcement learning, an agent acts in an environment, receives a reward signal and learns a policy: a strategy for choosing its next action.

  • 2016AlphaGo beat the world number one at Go, a feat thought to be decades away.
  • 2019OpenAI Five beat a professional Dota 2 team in real time.
  • TodayChatGPT’s RLHF uses human feedback as the reward signal for alignment.

Using AI responsibly

Session two covers six ethical challenges that anyone who builds or uses AI has to think about.

Bias and fairness

Models trained on biased data give biased outputs.

Facial recognition has been less accurate on darker skin tones.

Hallucinations

LLMs confidently generate false facts.

Always verify critical information.

Privacy

Training data often contains personal information.

Who owns your data?

Deepfakes

Synthetic media can deceive at scale.

Audio cloning, face-swapping and fake video.

Job disruption

AI automates both manual and knowledge work.

New roles appear in AI safety, prompt engineering and alignment.

AI safety

Alignment means AI doing what humans intend.

Safety teams at Anthropic, OpenAI and DeepMind work on it.

Design challenge: an AI for your school

Students pick a problem and answer four questions about it.

  1. What problem does it solve?Cafeteria wait times, lost library books, spotting bullying.
  2. What type of ML would you use?Supervised, unsupervised or reinforcement learning?
  3. What data would you need?How would you collect it, and who would label it?
  4. What could go wrong?Privacy, bias, who gets left out, and what if it’s wrong?

A worked example from class: school food waste

Predict how much food students will eat from past consumption, the menu, the weather and enrollment. Supervised learning fits the job.

The risks the class listed:

  • The model could be biased.
  • Cafeteria staff might rely on its predictions too much.
  • Student privacy suffers if it tracks what each person eats.

Where AI is in 2026

Students tour six areas of modern AI and the tools in each.

Large language models

GPT-4o, Claude, Gemini, Llama

Computer vision

YOLO, SAM, CLIP, GPT-4V

Voice AI

Whisper, ElevenLabs, Siri

AI coding assistants

Copilot, Cursor, Claude Code

Generative media

DALL·E 3, Midjourney, Sora, Runway

Autonomous systems

Tesla FSD, Waymo, Boston Dynamics

Tools to keep exploring

Build and experiment

Explore models

  • ChatGPT and ClaudeCompare responses from frontier LLMs
  • DALL·E and MidjourneyGenerate images from text prompts
  • WhisperTranscribe audio with near-human accuracy

Visualize and learn