You don't need a statistics degree to understand AI. You need about four ideas, held firmly.
The four ideas
Distribution. Don't ask "what's the value?" — ask "what's the spread of values?" Heights cluster near the middle; incomes pile up at the bottom with a long tail of outliers. The shape of the data decides which summaries are honest. Models learn distributions, not facts: a language model's entire skill is knowing the distribution of what word plausibly comes next.
Mean vs median. The mean adds everything and divides; the median is the middle value when sorted. The mean is dragged around by outliers, the median isn't. When a company says "average salary," reach for the median before you nod.
Variance. How spread out is the data? Two neighborhoods can share an average temperature — one a mild coast, the other a desert that swings from freezing to scorching. Same mean, wildly different experience. In AI, a model with high variance in its answers is unreliable even when it's right on average.
Sampling. You almost never see all the data — you see a sample, and samples can lie. Poll only landline owners and you've measured landline owners, not voters. Every dataset is a sample of the world, which means every model inherits the question: a sample of what, collected how?
Why this is the load-bearing floor
Every claim you'll ever read about AI — "95% accurate," "outperforms doctors," "state of the art" — is a statistical claim about a distribution, summarized into a headline number. The four ideas above are how you open that number back up and look inside.
What to read next
Statistics describes the data you have. Probability and uncertainty is about the data you don't — and it's the language models actually speak.