A low-stakes first guide
New to AI
You do not need to know technical language to begin. Start with a small question, keep private information out and treat every answer as something to check.
Estimated reading time · about 7 minutes
Keep This in Mind
AI can produce a useful response that is still incomplete, outdated or wrong.
Plain-language basics
What AI Can Help With
Useful Starting Points
- Suggest ideas or different ways to approach a question.
- Explain a topic in simpler language.
- Summarise text you provide.
- Help you compare options or find gaps in your draft.
What You Should Not Assume
- That an answer is accurate, current or unbiased.
- That the service is a private place for sensitive information.
- That confident wording means reliable evidence exists.
- That AI can take responsibility for your decision.
Choose one starting point
Choose One AI Tool to Begin
You do not need several accounts or the “best” tool before you can learn. Choose one official service, try a low-stakes question and focus on how you ask, check and decide.
OpenAI
ChatGPT
A general conversational AI service for questions, explanations and feedback.
Open ChatGPTGemini
Google's conversational AI service for questions, explanations and working through ideas.
Open GeminiMicrosoft
Copilot
Microsoft's conversational AI service for general questions, drafting and idea exploration.
Open CopilotAnthropic
Claude
Anthropic's conversational AI service for working through text, questions and explanations.
Open ClaudeStart with One
This is a neutral starting list, not a ranking, endorsement or claim that one tool is always better.
Features, prices, limits, sign-in requirements and privacy terms can change. Check the provider's current information, keep sensitive details out and, before a young person creates an account, ask an adult to review the provider's age requirements.
Do You Need to Choose a Model?
Usually not for a first test. Model names, availability and features change quickly. Begin with the standard chat experience offered by one provider, learn the basic habits and compare models later only when a real task gives you a reason. A newer, larger or paid model is not a substitute for checking important claims.
Your first use
Five Low-Stakes Steps
- Choose a small question
Use a topic where a mistake would not affect your health, safety, money, legal position or another person.
- Write what you already think
A sentence or two gives you something to compare with the AI response.
- Ask for help, not replacement
Request alternatives, questions or feedback instead of asking AI to make the whole decision for you.
- Check important claims
Use an appropriate, reliable source, especially for facts that may have changed.
- Compare and decide
Notice what changed in your view, what remains uncertain and what you will do next.
Safe example
Try a Prompt That Keeps You Involved
I'm learning about [topic]. Before answering, ask what I already think. Then give me two possible explanations, label anything uncertain and suggest one reliable source I could check. Do not ask for personal or sensitive information.
Replace [topic] with a low-stakes subject you want to explore. Read the answer critically rather than copying it as fact.
Output or action?
Did the AI Do It, or Only Describe It?
AI can generate text in the shape of a terminal log, search summary, citation, data response or confirmation message. A realistic-looking answer is not proof that a tool ran or the claimed action happened. Unless the service actually used a connected tool and you can inspect the result, the answer does not establish that anything was searched, calculated, sent, saved or changed.
Some AI services can browse, run code or use connected tools. Others cannot, and a tool-enabled service may not use a tool on every turn.
What an Answer Can Imitate
- A command-line or app response.
- A web-search summary or citation.
- Structured data or a success message.
- Confident wording that says the task is complete.
What You Should Verify
- Was a real tool available and actually used?
- Can you open the named source or destination?
- Is there a receipt, changed file, sent item or live result?
- Does the result you can see match what the AI said?
Check the Result, Not Just the Message
Look for the receipt, sent item, changed file, source page or live result. A genuine tool call can still use stale input, follow the wrong instruction or return an error.
If you cannot inspect the result, say that you are not sure. The AI repeating that it worked is not a separate check.
Use an honest status: Verified result, Description only, Evidence conflicts or Not sure. Even with a verified result, review it before relying on it.
Why this check was added
In a 2022 article, Jonas Degrave documented an experiment in which ChatGPT was asked to imitate a computer terminal. It produced convincing network output from an imagined system rather than a real connection. Andrej Karpathy later discussed the example at TreeHacks 2023. This historical example shows why a realistic-looking response should be checked. It does not prove that every current AI service works the same way or that this guide improves learning.
Privacy and high-stakes boundaries
Know When to Stop and Verify Elsewhere
Keep Sensitive Details Out
Do not enter login details, addresses, financial details, health records, confidential workplace information or identifying information about another person.
For medical, legal, financial or safety questions, use qualified help and current official sources. Do not rely on an AI answer alone. In an emergency, contact the relevant local emergency service rather than relying on AI.
Practise next
Check Your Judgement
Use the Thinking Check for broad scenarios, practise evidence matching in the Claim Check Lab, or apply the fuller exercise to a real low-stakes question.