AI can write essays, paint portraits, and even chat with you like a friend. So why does it struggle with something as simple as drawing a person writing with their left hand?

Turns out, the answer reveals everything about how AI learns—and how it fails.

1. AI Isn’t Magic. It’s Math.

First, let’s bust a myth: AI doesn’t think like a human. It doesn’t imagine, reason, or wonder. It doesn’t “know” what a left-handed person is.

Instead, AI works through algorithms—sets of math instructions that find patterns and make predictions. You feed it examples, and it looks for trends. That’s it.

If it never sees a left-handed writer in its training data, it won’t know how to generate one. That’s not stupidity—it’s statistical blindness.

2. Models Learn from Data. Period.

Think of it this way:

The algorithm is the recipe.
The AI model is the cake.
The data is the ingredients.

Even with the same recipe, two cakes can turn out totally different if the ingredients are off. If 99% of the training images show right-handed people writing, the model will assume right-handedness is the rule. So when it’s asked to “draw someone writing,” it defaults to what it knows: right-handed images. Why? Because the model has adjusted its weights (what matters) and bias (its built-in lean) to fit the training data.

3. Wait—What Is “Bias” in AI?

Let’s clear this up: AI bias ≠ human bias… but they’re deeply connected.

In AI, “bias” is a mathematical term in the formula:

z=w1x1+w2x2+⋯+wnxn+b

That b is the bias—just a number that shifts the output up or down. It helps the model make decisions, even when inputs are zero.

But here's the twist: psychological bias can cause biased AI models, because…

The model’s “bias” terms are adjusted during training — and if the training data reflects cognitive bias (like overrepresentation of one group), the model will mathematically learn that lean. (like stock photos of mostly white, right-handed men), the bias term in the math starts to reflect human cognitive bias.

Example: Imagine an AI trained to predict job candidates’ success.

If the data shows that most past hires were male (due to past human bias),
The model will assign higher weights to features that correlate with male candidates,
And the bias parameter shifts the threshold accordingly.

So: Cognitive bias in data leads to biased model outputs, and this is encoded into the model’s weights and bias parameters.

4. How Machines “Learn”: The Three Ways

AI doesn’t go to school—but it still learns. Let’s break down how:

Supervised Learning:

AI is trained with labeled data.
Example: "This is a cat." "This is a dog."
If no one labels “left-handed person,” AI misses it.

Unsupervised Learning:

AI finds patterns without labels.
Example: Clustering people by how they walk.
But if left-handedness isn’t a strong signal in the data, it’s overlooked.

Reinforcement Learning:

AI tries things, gets feedback (rewards), and improves.
Example: Learning to play chess or walk.
But again—it only improves on what humans choose to reward.

So what happens if no one highlights left-handed examples, or values them as important? AI just… doesn’t learn them.

5. Garbage In = Garbage Out

Training AI is like prepping a student for an exam using flashcards. If the flashcards never mention “left-handed posture,” the student flunks that question—no matter how smart they are.

A chatbot might:

learn grammar using self-supervised learning (a twist on supervised)
be fine-tuned by humans (supervised)
and then get better through feedback (reinforcement learning)

The Training Process: How a Model is Built

Here’s what happens when you train an AI model:

Choose a model structure — What kind of model fits your problem?
Collect data — Lots of examples are needed to learn from.
Clean and label it — Fix errors, organize it, and label (if needed).
Pick your hyperparameters — These are training settings like learning speed.
Train the model — Feed it data, calculate how wrong it is, and adjust it.
Optimize — Use gradient descent or another method to keep improving.
Test it — See how it performs on new data it’s never seen before.

This cycle repeats over and over until the model gets accurate enough!

If an AI draws 6 fingers or forgets lefties, it’s not because it’s “dumb.” It’s because no one ever taught it properly.

And this opens up big questions:

What else has AI never seen?
What assumptions is it quietly making?
What does that mean for fairness in hiring, policing, or healthcare?

Because behind every biased output is a human training choice—sometimes invisible, sometimes unintended.

Fun Fact

In 2016, Microsoft launched Tay, a chatbot designed to learn how to talk by chatting with humans on Twitter. But within 16 hours, Tay turned racist, offensive, and had to be shut down.

Why? It was trained in real time—on unfiltered tweets from the internet. The model had no safeguards, so it learned exactly what people fed it.

Lesson? If your training data is toxic, your AI becomes toxic. It's not evil. It's just... under-taught.