Zulia Shavaeva← Back to Zulia
Free AI and security guide

Free learning route

AI fundamentals

Learn what machine learning and large language models are, why their answers can be wrong, how to write a clear prompt and how to check an AI answer. Use the examples on this page with a pen or notes file.

For: Anyone who keeps hearing about AI at work and wants to understand it properly, with no technical or coding background.

Before you start: None. No coding, no maths, no account, no software to install.

First session: about 30 minutes. Times are estimates; work at your own pace.

Try the first taskChoose a different route

Tick completed steps to save progress in this browser. No account or quiz needed. Clearing browser data removes progress.

What you will be able to do

  • Explain in your own words the difference between traditional software, machine learning and a large language model
  • Recognise four common problems in AI answers and explain what evidence would help you check them
  • Write a prompt that states a task, input, format and limits
  • Review an AI-generated answer against a short checklist and record what you would verify before using it

Start now · about 20 minutes

Grade three AI answers for a fictional company

  1. Copy this fictional scenario into a notes file: Northwind Bikes, a made-up bicycle retailer, asks an AI assistant 'How many refunds did we issue in March and what caused them?'
  2. Write out the three fictional answers given below as answer A, B and C. A: 'Northwind Bikes issued 1,248 refunds in March, mostly for brake faults.' B: 'I do not have access to your refund data. To answer this you would need the March refund export, and I can then group the reasons for you.' C: 'Refunds typically rise in spring due to weather. Your March figure is likely in the normal range.'
  3. For each answer write one line saying whether the assistant had the data it needed, and one line saying what a reader might wrongly believe after reading it.
  4. Rank the three answers from most to least trustworthy and write one sentence explaining your top choice.
  5. Write the single question you would ask Northwind Bikes before letting any of these answers go to a customer.

Done when: You have a short note ranking A, B and C, with answer B ranked first because it states its own limits, and one verification question written down.

Lesson 1 · about 30 minutes

What people actually mean when they say AI

Traditional software often applies rules people write. For example: if a basket total is over £50, offer free delivery. Machine learning instead learns patterns from examples. A model is the resulting system that maps inputs to predictions; training is the process of learning those patterns. People still choose the data, design the system and decide how to test it.

A model can reflect gaps and biases in its training examples. A loan model trained on past decisions may repeat unfair patterns in those decisions. Good performance on old examples does not guarantee good decisions for a different group of people or a different setting.

A large language model, or LLM, learns patterns in language. Many LLMs begin training by predicting the next token, a small unit of text, and are then trained further to follow instructions. Applications may also add search, tools and checks. LLMs can draft, summarise and explain, but fluent text is not evidence that an answer is true.

Generative AI produces content such as text, images or code. An AI agent uses a model to help choose and carry out steps, often through software tools. A drafting assistant may only suggest an email; an agent with a sending tool can actually send it. Those permissions change the consequences of an error.

Try it

Take three fictional tasks at Northwind Bikes: (1) calculate VAT on an order, (2) predict which customers are likely to return a bike, (3) draft a reply to a complaint email. Label each as ordinary software, machine learning, or large language model, and write one line of justification.

Check your answer

One reasonable choice is ordinary software for VAT, because the calculation follows an explicit rule; machine learning for a return prediction, because past orders may reveal useful patterns; and an LLM for drafting a reply, where several wordings may be acceptable. These are design choices, not rigid categories. Prefer a simpler, testable method when it meets the need.

Read more

Lesson 2 · about 30 minutes

Why a confident answer can still be wrong

Here are four common problems to watch for, not an exhaustive list. First, unsupported or invented claims, often called hallucinations: a model may give a statistic, case name or product feature that the evidence does not support. Second, stale information: training data may be out of date, and retrieved sources can be old too.

Third, missing context. An assistant cannot reliably report your March refunds unless the application gives it the relevant records. It may ask for those records or say it cannot answer, but it can also guess. Fourth, misleading confidence: fluent wording or a confident tone does not tell you whether the answer is well supported.

Grounding means providing source material, such as a document or database result, to support an answer. Retrieval augmented generation, or RAG, retrieves relevant material and supplies it to the model with the question. This can improve answers, but retrieval can miss a source and the model can misread it. Open citations and check that they support the claim.

Treat an AI output as a draft that needs a level of checking appropriate to its use. A brainstorm can tolerate errors that a payment, legal statement or customer commitment cannot. Check important claims against authoritative sources, and use an appropriate specialist where the decision needs one.

Try it

Here is a fictional AI answer for Northwind Bikes: 'Under the UK Cycle Retail Act 2019, refunds must be issued within 7 days. Northwind's average is 9 days, so you are non-compliant and were fined last year.' Write down each separate claim on its own line and label it with the failure mode it most likely shows, then write what you would need to see to accept it.

Check your answer

The legislation, refund average and fine are three separate claims. In this fictional example, the Act and fine are invented and the average has no supporting dataset. Do not infer the cause of a real error just from its wording. Verify legislation with an official source, an average with the relevant records and a fine with an actual enforcement notice.

Read more

Lesson 3 · about 25 minutes

Writing a prompt that gets a usable answer

A prompt is the instruction and context you provide to a model. A useful starting structure is: task, input, output format and limits. An optional role can help set the perspective, but it does not give the model real qualifications. Clear instructions improve the chance of a useful answer; they do not guarantee one.

Compare two prompts for the fictional Northwind Bikes. Vague: 'Write something about our refund problem.' More specific: 'Using only the refund notes below, list the three most frequent refund causes. Output a table with cause, count and one example note ID. If the notes do not support a ranking, say so instead of guessing.'

Asking the model to report missing information makes the expected behaviour clear. Asking for a table can make gaps easier to see. You still need to check the counts and source IDs: instructions such as 'use only these notes' are not a technical guarantee that the model will comply.

Keep a simple prompt log: the prompt, model or tool, date, what you changed and what you checked. Use fictional information for practice. At work, use only approved tools with data you are permitted to share, following your organisation's rules and the tool's data-handling terms.

Try it

Rewrite this weak fictional prompt so it names role, task, input, format and limits: 'Summarise our customer feedback and tell us what to fix.' Keep it under 80 words and include one instruction about what to do when the input is insufficient.

Check your answer

One example: 'Using only the 20 fictional feedback comments below, group them into at most five themes. Return a table with theme, count and a quoted example. Rank by count. Do not suggest fixes. Put themes with fewer than three comments under Other. If a comment is too vague to group, list it separately.' Check that your prompt defines the task, permitted input, format and treatment of uncertainty. Then verify the output rather than assuming the prompt worked.

Read more

Keep what you made

One-page AI working note

One page containing your own definitions of machine learning, LLM, generative AI and grounding; four common answer problems with examples; a prompt template; and three checks you would make before using an AI output. Keep all examples fictional.

  • Every definition is in your own words, with no phrase copied from a source
  • Each of the four failure modes has a concrete example a colleague would recognise
  • The prompt template names the task, input, format and limits, including what to do when information is missing
  • Your three questions include at least one about where the data came from
  • No real customer, employee or company data appears anywhere on the page

Completing a route records your practice, not a qualification or proof of job readiness.

Where to go next

Back to all learning routes