Train the Bot
Tap creatures to pick examples, then Train
What is Train the Bot?
Train the Bot is a free game that shows how machine learning actually works, and where it goes wrong. Creatures are either Blips or Blops according to a hidden rule. You choose which examples the bot gets to study, it works out its own rule from only those, and then it is tested on creatures it has never seen. Pick a one-sided set and the bot confidently learns the wrong thing - which is exactly how bias gets into real AI systems.
How to play Train the Bot
Controls: Tap creatures to pick examples, then Train
- 1You are shown twelve creatures, each labelled with its true answer: Blip or Blop.
- 2Tap creatures to add them to the bot's training set. You only get a few picks.
- 3Press Train the bot. It looks at your examples and works out a rule by itself.
- 4The bot's rule is shown in plain English, then it is tested on eight creatures it has never seen.
- 5Your score is how many it gets right. Five rounds per game.
Train the Bot tips and strategy
- Include both answers. A training set of only Blips teaches the bot nothing, because there is no difference for it to find.
- Vary everything that does not matter. If every Blip you pick happens to be blue, the bot will decide Blips are blue.
- Look for pairs that differ in one feature only. Two creatures that are identical except for their spots tell the bot far more than two completely different ones.
- When the bot gets it wrong, read the rule it invented. It is always a real pattern in your examples - just not the one you meant.
Train the Bot FAQ
How does an AI actually learn?
It looks for a pattern that separates the examples it was given, then applies that pattern to new things. It does not understand what any of it means. This game makes that literal: the bot tests every simple rule it can form and keeps whichever one fits your training examples best.
What is AI bias, in simple terms?
Bias is when the examples an AI learned from were one-sided, so the pattern it found is not the pattern you wanted. If every Blip you showed it was blue, it decides Blips are blue and then fails on a red one. Real systems fail the same way when their training data does not represent everyone.
What is training data?
The examples an AI learns from, each labelled with the right answer. In this game, the creatures you pick are the training data, and the eight creatures at the end are the test data - deliberately kept separate, exactly as real machine learning does it.
Is this how real AI works?
It is a genuine, very simplified version. Real classifiers weigh many features at once rather than picking a single rule, but the core loop - learn a pattern from labelled examples, then predict on unseen data - is the same, and so is the failure when the examples are skewed.
