AI fluency test

What is the best description of a hallucination?

A response the model refuses to give for safety reasons or because the topic is outside what it was built to discussA fluent, confident answer that contains made-up or wrong factsA response that changes each time you ask, for example one answer with a detail and the next one without itA response that ignores part of your question, often because the prompt was too long or had several questions in one

Why do AI models sometimes get letter-counting questions wrong?

They do not understand numbersTheir training skipped short wordsThey only read the first half of each wordThey read text in chunks called tokens, not in single letters

A chatbot answers a question about last week's news incorrectly. What is the most likely reason?

Its training data ends at a cutoff date, and it has no search tool to look things upIt searched, but the search results were wrong or off-topic, and the model repeated them without checking whether the pages were reliableIt avoids recent events on purpose, for safetyIt needs a longer prompt to remember what happened

What does the 'context window' limit?

How long the AI takes to respondHow many people can use the model at the same timeHow much text, including your prompt, the chat history, and the answer, the model can handle in one goHow many facts the model knows in total

You rerun the same prompt and get a different answer. What explains it?

The model learned from your first attempt, so the second answer reflects what it picked up from the first oneThe model picks among likely next words, so some randomness is built inThe provider quietly switched your account to another model version, which happens when a newer release replaces the one you used earlier todayThe first answer was cached and then expired, so the system generated a new one the next time

You want help writing a cover letter. Which opening prompt will probably give the best first draft?

Write a great cover letter for a marketing jobWrite a cover letter. Make it professional and impressive, use strong action verbs, avoid cliches, and keep the formatting clean and polishedHere is the job post and my CV. Write a 250-word cover letter that links my experience to the three main requirementsWrite a 250-word cover letter for a marketing job in a confident tone

The AI's answer is good, but the tone is too formal. Best follow-up?

Keep the content, but rewrite it in a friendly, casual tone, like talking to a coworkerMake it betterToo formalStart again with a different tone, and this time list every point you want to keep so nothing from the first draft is lost

Why does asking a model to 'think step by step' often help on maths and logic questions?

It switches on a built-in calculatorIt slows the model down on purpose so it checks moreIt makes the answer longer, and length alone improves accuracyWriting out the steps gives the model intermediate results to build on, so it makes fewer jumps

You need the AI to follow a strict format every time. What works best?

Describe the format in a long paragraph, listing each field, its type, and the order it should appear in, so nothing is left openShow one or two examples of the exact format and say 'return only this format'Ask it to be very careful with formattingAsk for the format at the end of a long chat

You paste a long article and ask for a summary, but it misses the key point. Best fix?

Ask again with the same promptPaste the article twiceAsk for a longer summary, so the model has room to include the points it skipped, and ask for headings to organize themSay what the key point should cover and who the summary is for, and ask it to quote the lines it relied on

The AI says: 'Einstein failed math in school.' You want to use this in a presentation. What do you do?

Use it, it is a famous factUse it, but say 'some say' and add a note that the claim is widely repeated, so the audience knows you are not certainCheck a reliable biography or encyclopedia first, because it is a popular mythAsk the AI for sources and use those

The AI writes a legal clause and says it is 'valid in all US states'. Best response?

Treat it as a draft and have a lawyer check it, since laws differ by state and the AI does not know your caseUse it, the AI was specificAsk the AI to double-check it state by stateCompare the wording with a template you found online

The AI backs a claim with: 'Smith et al., 2021, Journal of Applied Cognition, pp. 112-130'. Best next step?

Trust it, it has page numbersSearch for the paper title only, and if a paper with that title exists, treat the citation as confirmed and move onAsk the AI if the citation is realLook the paper up in Google Scholar or a library database and check that it says what the AI claims

You ask the AI for the average of 14 numbers and it gives a confident answer. What is the sensible check?

None, computers are good at mathsRecalculate it with a calculator or spreadsheet, since language models can make arithmetic slipsAsk it to explain its answer step by step, then check that every step follows from the one before it and the totals add upAsk the same question again and see if the answer stays the same

The AI summarizes your meeting notes and says 'Priya will send the budget by Friday'. Your notes never mention a deadline. What happened?

It filled an obvious gap, so the summary is fineIt made a reasonable guess, which is fine for a draft as long as you tell the attendees it is a draft and ask them to correct anything that is offIt added a detail that was not in your notes, so check each action item against the sourceThe deadline was probably in its training data

Which is the best habit before pasting text into a chatbot?

Remove names, account numbers, and anything you would not post publicly, or use a tool your company approvesCheck that the page has a lock iconDelete the chat afterwardsTurn on private browsing

A free 'AI PDF reader' site asks you to upload your tax return to get a summary. What is the sensible decision?

Upload it, summaries are harmlessUpload only the first page, since that page usually has fewer personal details than the rest of the returnUpload it if the site has good reviewsDo not. Use a tool you trust with a clear privacy policy, or summarize it by hand

Your teacher allows AI for brainstorming but not for writing. You ask AI for an outline, then write the essay yourself. What is the safest approach?

That is fine, because the AI only helped you think and you wrote every sentence of the essay yourself, which is what the rule is really aboutRe-read the rules, ask the teacher if you are unsure, and say how you used the AISay nothing, since the writing is yoursAsk the AI to rewrite your essay so it cannot be detected

An AI coding assistant puts your database password directly in the source file. What do you do?

Leave it, the repo is private, and add the file to your notes so you remember to move it before the project goes publicReplace the password with 'xxxx' before sharing screenshotsMove it to an environment variable, keep it out of git, and change the password if it was ever committedPut the file in a zip with a password

Who is responsible for a mistake in an email that an AI wrote and you sent?

You are, because you chose to send itThe AI companyBoth of you, equallyNobody, if you say 'AI helped'

AI Basics

Concepts

Prompting

Spot the mistake

Safe use

AI Basics

Concepts

Prompting

Spot the mistake

Safe use