How to learn AI with no technical background: the first week, two skills and one safety rule

Learning AI as a user is not learning to build it

People who type how to learn AI into a search box often expect a syllabus of mathematics and code, the road of engineers who build models. A user needs something else: to set a task for a ready-made assistant and judge what comes back. UNESCO's AI competency framework for students, published in 2024, shows the proportion: 12 competencies in four dimensions, a human-centred mindset, ethics of AI, AI techniques and applications, and AI system design. Only the last is about designing systems.

The same framework sets three levels: understand, apply, create. The AILit Framework, a joint initiative of the European Commission and the OECD, orders its four domains similarly: engage with AI, create with AI, manage AI, shape AI. Both were written for schools, yet the order suits an adult beginner: use the tool with judgement first, build systems later, if at all. The starting kit is small: one assistant, your own tasks and the habit of checking.

How a language model works, in three honest sentences

Three sentences are enough to begin. A language model estimates how likely each next piece of text is, given the text before it; Google's machine learning guide for developers calls the pieces tokens: a word, part of a word or a single character. A large model repeats this at scale, and the guide describes such models as essentially autocomplete mechanisms that can predict thousands of tokens. The result reads like knowledge but remains a prediction: these models, the guide says, hallucinate, meaning their predictions often contain mistakes.

The first consequence: the model can be wrong with full confidence. Anthropic's help centre says its own assistant can produce quotes which may look authoritative or sound convincing but are not grounded in fact, and may lack up-to-date training data. Its advice: do not treat the assistant as a single source of truth, and scrutinise any high-stakes advice. A fluent tone is a property of the text, not evidence that the text is right.

The second consequence: one request does not always return one answer. OpenAI's prompt engineering guide states that generated content is non-deterministic. Different types of models, it warns, may need different prompting, and even snapshots within one family can produce different results. So lists of magic phrases age with every release, while a clear task and a checked result carry over to any assistant.

The first week: one assistant, your own tasks and a notebook

Choose one assistant and stay with it for seven days. Comparing five tools on the first evening teaches you their menus, not the skill. Bring it work you would do anyway: a reply to a difficult email, a long agreement to summarise, a messy list to turn into a plan. Real tasks have a real standard, because you already know what a good result looks like.

Keep a notebook with five short lines for every attempt: the task, what you asked, what came back, what you changed, and whether you would use the result. After a week there are a dozen entries or more, and they show where the assistant saves you an hour and where it costs you one. Add one habit: each day, ask something whose answer you already know. Nothing calibrates trust faster than watching the tool slip on your own ground.

The skill of asking: role, context, limits and the form of the answer

The companies that make the models describe a good request in nearly the same words. Anthropic's prompting guide suggests thinking of the model as a brilliant but new employee who lacks context on your norms and workflows. Its golden rule is a test anyone can run: show the request to a colleague who knows little about the task. If they would be confused, the model will be too.

In practice a request has four parts. OpenAI's guide arranges an instruction in sections: identity, instructions, examples and context. Google's prompting guide for Gemini adds constraints and the response format: a table, a bulleted list, keywords, a sentence or a paragraph. In everyday terms: who is answering, what the situation is, what is off limits and what shape the answer takes. Anthropic notes that even one sentence of role makes a difference, as does explaining why an instruction matters.

Two more habits complete the skill. The first is examples: Google's guide recommends always including a few, and Anthropic advises three to five, relevant and varied. The second is patience with the first draft: Google writes that prompt design can take a few iterations and suggests rephrasing the request or reordering its parts. Rather than open a new chat and ask again as vaguely, stay in the conversation and say what was wrong.

The skill of checking: sources, dates and numbers

Asking well is half the work; the other half begins when the answer arrives. Anthropic's documentation on reducing hallucinations opens with an admission: even the most advanced language models can sometimes generate text that is factually incorrect or inconsistent with the given context. Its techniques lower the risk and, by its own words, do not remove it. The closing line is the rule to keep: always validate critical information, especially for high-stakes decisions.

Three things deserve a check every time. Sources: open each one, because a quotation may be invented, and Anthropic's help article adds that the original site can hold context missing from the summary. Dates: training data ends at some point, and Google's prompting guide advises switching on search grounding whenever the model may need obscure or recent facts. Numbers: recalculate them, since its machine learning guide observes that a model can give the appearance of sophisticated mathematical reasoning while only completing a word problem.

Part of the checking can be handed to the model itself. Anthropic suggests giving it explicit permission to say it does not know, asking for word-for-word quotes from a long document before any conclusions, and having it withdraw every claim it cannot back with a quote. Running one request several times also helps: disagreement between versions may point to an invention. None of this replaces your own eyes on anything that touches money, health or law.

The next step: documents, tables, images and simple automation

After a week of plain conversation, start giving the assistant material. For long documents Anthropic offers a concrete rule: put the document at the top and the question at the end; in its tests this improved response quality by up to 30 percent. With a table, have the columns described before anything is calculated, and compare one total with your own count. For images, Google's image generation documentation advises describing a scene in rich detail: the more specific you are, the more control you have.

Automation starts earlier than most people think. When you type the same request for the third time, save it as a template with the four parts filled in and blanks for what changes. The next level is a chain of steps; Anthropic calls self-correction the most common pattern: a draft, a review against criteria, a refinement. Agents grow out of this, but a small library of templates is already automation and needs no code.

Safety with an AI assistant: what not to paste into a chat

Every provider publishes a notice on what happens to your conversations; read it before the first serious task. Google's Gemini Apps Privacy Hub asks users not to enter confidential information they would not want a reviewer to see. A subset of chats, it explains, is read by human reviewers, and reviewed chats are kept for up to three years even if you delete your activity. By default, it adds, activity is auto-deleted after 18 months; the period can be changed to 3 or 36 months.

Anthropic's privacy centre describes another arrangement for its consumer plans: chats may be used to improve the models when the user allows it in the settings, incognito chats are not used, and rating a reply stores the whole conversation for up to five years. Terms differ by provider and plan and change, so read the current text. The working rule is simpler: no passwords, no document numbers, no other people's personal data, no client files. Where a contract or a data protection law covers your work, ask a specialist first.

What you can practise alone and what a structured course adds

Everything above can be practised alone, starting this week: one assistant, five tasks of your own, the notebook, the four-part request and the three checks. It costs attention and nothing else. What self-study gives less readily is order and feedback. Alone, people repeat what already works and skip what is uncomfortable, usually verification and safety.

A structured course supplies that order. The Artificial Intelligence course at VITON13 SCHOOL is taught online, with practice on the learner's own task. Its programme has seven modules: working with AI deliberately, prompting as engineering, research and verification, automating processes, AI agents, your own systems, safety and responsibility. Current terms are on the course page. If your notebook shows the same walls week after week, a programme saves time; otherwise, practise first. The question of how to learn AI then gives way to a better one: which task to hand over first.