The more varied and representative the data, the better the AI generalises to new situations. A model trained only on photos of golden retrievers might fail to recognise a poodle. Diversity in data leads to robustness in AI.

This is why companies invest heavily in collecting, cleaning, and curating large datasets - it is often the most time-consuming and expensive part of building an AI system.

The good news is that once a high-quality dataset exists, it can be reused and shared, accelerating research worldwide.

Data Quality: Garbage In, Garbage Out

AI is only as good as the data it learns from. If the data is messy, incomplete, or incorrect, the AI will produce unreliable results. This principle is known as garbage in, garbage out (GIGO).

Common data quality problems include:

• Missing values - blank fields that leave gaps in what the AI can learn.
• Duplicates - the same example appearing multiple times, skewing the model.
• Incorrect labels - a photo of a dog labelled as a cat confuses the learning process.
• Outdated information - training on data from 2005 will not reflect 2025 trends.

Real-World Datasets

Some datasets have become famous in the AI community:

| Dataset | What It Contains | Size | |---------|-----------------|------| | ImageNet | Labelled photographs | 14 million+ images | | Common Crawl | Web pages from across the internet | Petabytes of text | | Wikipedia | Encyclopaedia articles | 60 million+ articles | | MNIST | Handwritten digits (0–9) | 70,000 images |

These datasets are freely available and have been used to train some of the most influential AI models in history.

Data Labelling and Annotation

Before AI can learn from data, someone usually needs to label it. This means telling the system what each example represents.

• A photo gets tagged as "cat" or "dog."
• A sentence gets marked as "positive sentiment" or "negative sentiment."
• A medical scan gets annotated by a doctor as "healthy" or "abnormal."

This process is called annotation, and it is often done by humans - sometimes thousands of them working together.

Data Bias: When Data Gets It Wrong

Data reflects the world it comes from - including the world's prejudices. When a dataset over-represents one group or under-represents another, the AI trained on it will inherit those imbalances.

Real examples of data bias:

• A hiring algorithm trained mostly on male CVs learned to penalise female applicants.
• Facial recognition systems trained mostly on lighter skin tones performed poorly on darker skin tones.
• A medical AI trained on data from one country missed symptoms common in other populations.

Key Takeaways

• A dataset is an organised collection of examples that AI learns from.
• Data can be structured (tables) or unstructured (images, text, audio).
• Data quality directly determines AI quality - garbage in, garbage out.
• Labelling is the human effort that teaches AI what each example means.
• Bias in data leads to bias in AI - and it has real-world consequences.

Next up, we will look at the algorithms that actually process all this data and turn it into intelligence.