Act 1 · Lesson 2 of 21≈6 minutes○ Saves on this device

A Diet of Words

Discover what AI was fed — and why its diet explains both its brilliance and its blind spots.

By the end you'll be able to

  • Describe training data as the collection of examples that patterns are learned from.
  • Explain why gaps or imbalances in the examples become gaps or imbalances in behaviour.
  • Spot the kinds of questions a text-trained model is poorly equipped to answer.
⚡ 30-second warm-up — From Lesson 1: your colleague's AI gave a confident grant deadline. Why couldn't you trust it?

Because generation is not verification — the date was a likely continuation of patterns, not a looked-up fact. Today: where those patterns come from.

Revisit Lesson 1

Start with a thought experiment

Imagine you learned English entirely from cookbooks. Thousands of them — nothing else, ever.

You'd be superb at some things and hopeless at others. Which?

Commit to a prediction

The cookbook-educated speaker is asked to write a sympathy card. What happens?

This lesson changes one thing: you'll stop thinking "the AI knows things" and start thinking "it reflects a diet of examples — thick in places, thin in others — and its tone won't tell me which I'm getting."

The one idea in this lesson

How it works Core concept

Lesson 1's patterns have to come from somewhere. Builders assemble enormous collections of text — websites, books, code, conversations — and use them as examples. This collection is called training data, and the model's behaviour is a compressed reflection of it: nothing in, nothing out; skew in, skew out.

The collection is not the whole of human knowledge. It over-represents English, the internet-connected, the recent past and the frequently written-about. It under-represents private, local, very new and rarely written-down things.

And here is the part that matters at work: expressed confidence is not reliably calibrated to the thickness of the evidence. The model continues patterns from thin evidence as fluently as from thick — and neither a confident nor a cautious tone dependably signals which you're getting. Asking about something obscure is asking the cookbook reader to write that sympathy card — you'll get something, and it will sound composed.

frontier assistants patch some diet gaps by fetching live pages and your own documents at answer time (retrieval — Lesson 13 shows exactly how, and what new checking it demands). The diet remains the foundation: it shapes how the engine writes about whatever gets fetched, and everything answered without a fetch comes straight from it.
The diet, to scale Horizontal bars show the rough make-up of a training diet: web text is by far the widest bar, then books and reference, code and technical text, then conversations and forums. Beneath them, three barely visible slivers are labelled your suburb's news, niche funders' rules, and anything after the collection date — showing that the confident voice is built mostly from the big bars, not the slivers. Web text Books & reference Code & technical Conversations your suburb's news · a niche funder's rules · anything after the collection date Illustrative proportions — exact mixes vary by provider and are rarely fully disclosed.
The confident voice is built almost entirely from the big bars. When your question lives in the red sliver, the voice doesn't change — the evidence underneath it does.

A picture to hold on to

Useful mental model Analogy

A chef's palate formed entirely by the restaurants they've eaten in. Eat widely, and their range is extraordinary. But if whole cuisines were missing from their life, strange gaps appear — and every new dish they invent can only be a variation on what they've tasted.

  • Restaurants visited = the training examples.
  • The palate = the learned patterns.
  • Cooking a new dish = generating an answer.
  • Missing cuisines = under-represented topics, languages and communities.

Test the diet yourself

Conceptual illustration — not measurements from an actual model. Pick a topic and predict first: thick coverage (written about constantly) or thin (rare, local, private or too new)? Then reveal the evidence gauge — and read the simulated answer's tone carefully. (Nothing is sent anywhere.)

Where this bites at work

In the field · Luka

Luka, a nonprofit fundraiser and grant acquittal consultant, asks about the acquittal requirements of a niche philanthropic trust. General grant-reporting patterns are thick in the training diet; this particular funder is vanishingly thin. The answer blends both seamlessly — correct-sounding general practice, wearing the funder's name.

His habit now: before relying on an answer, he asks, "Would the internet have written much about this, before roughly the model's collection date?" For a niche funder's rules — no. So he goes to the funder's own guidelines.

Where the picture breaks

Analogy limits

A chef remembers individual meals and can tell you about the restaurant. The model cannot: it doesn't store its examples as retrievable documents. What remains after training is distilled tendency — the way your old maths homework is gone but your arithmetic remains.

A chef also chooses what to eat. A model's diet is chosen by its builders — which means the gaps and skews are human decisions with human accountability, a thread we pick up properly in Lesson 18.

Now make the call

Luka's assistant has just produced the acquittal requirements for the Harwood Trust — "reports due within 60 days, income and expenditure statement signed by the treasurer." The acquittal must be drafted today, and his client is a volunteer committee that trusts him. What does he do?

Apply it to your work

This week, at your desk

List three topics from your own work that the training diet is almost certainly thin on — your funders' specific rules, internal policies, local programme detail. Write them on your fact-check checklist as permanent verify-first zones.

Print the AI fact-check checklist

Your habit from this lesson

Do this from today

Before relying on an AI answer, run the diet question: "Would the internet, before roughly the model's collection date, have written much about this?" Thick topic — proceed with normal care. Thin topic — treat the answer as a well-dressed guess and go to the source.

Go deeper (optional): what goes into the diet, and who decides

Assembling training data is active curation: deduplicating near-copies, filtering low-quality and harmful text, weighing sources, handling licensing, and increasingly adding synthetic examples — text generated by other models. Each choice shifts the resulting behaviour; none of it is neutral, and providers differ in what they include and disclose.

Every model also has a collection cut-off: a date after which the world simply isn't in the diet. Products bolt on live search to compensate — but that's retrieval (Lesson 13), not diet.

Deeper challenge: list three topics from your own job that you'd bet are thin in any model's diet. (Internal policies? Local suppliers? Anything behind a login?) Those are your permanent verify-first zones.

Prove it to yourself

Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct). This is what defeats the feeling of "that made sense" evaporating by next week.

Your field note

One or two sentences in your own words: why does an AI answer some questions brilliantly and others badly — in the same confident voice?

  • Mentions the diet — the examples it was trained on
  • Distinguishes thick coverage from thin
  • Notes that tone doesn't signal which — so check thin topics

Evidence & review — how we know what this lesson claims
Claim register for Lesson 2
ClaimTypeBasisReview risk
Training corpora over-represent English, online and frequently written-about materialHow it works (empirical)Corpus research literatureLow
Expressed confidence is not reliably calibrated to evidence thicknessHow it works (empirical)Calibration researchMedium
Modern assistants patch diet gaps with live retrievalProduct behaviourProvider documentationHigh — reviewed quarterly
The diet-proportions figureTeaching deviceIllustrative, labelled; exact mixes rarely disclosedMedium
The chef's-palate / cookbook analogyUseful mental modelLimits stated in the lessonLow

Concept review due: January 2027. Product-behaviour claims: quarterly.

Where you are in the machine

You're in the foundations: the training-data layer. Select any layer for its role, lessons and your live progress.

The machine, bottom to top: training data (this lesson) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).

Next: what does "reading" even mean to a machine?

The machine "eats text" — but computers only eat numbers. Before any pattern can be learned, language must be chopped into pieces, and the chopping explains some famously silly AI failures. Lesson 3: Chopping Language into Lego.