Thinking Check

You want help wording a message about a real customer. What should you do first? Paste the full message with the customer's name so AI has enough context, then ask it to remove personal details from the final draft. Replace identifying details with fictional or generalised information and check the relevant privacy or workplace rules before using anything real. Use the customer's initials in a private browsing window, but keep their dates and situation unchanged so the advice stays specific.

AI gives you a surprising fact for an assignment. What is the strongest next step? Use it if the answer is detailed, confident and fits what I already know about the topic. Ask the AI the same question again in a new chat and accept the fact if the second answer agrees. Find a suitable reliable source, then check the fact's wording, date and context before using it.

You are learning to write a cover letter. Which approach best keeps authorship with you? Write my key points and first draft, then ask AI to identify unclear parts and suggest questions for me to consider without rewriting it. Ask AI for three complete versions, choose the one that sounds most like me and adjust a few words before sending it. Give AI a successful example from the internet and ask it to adapt the wording to my work history.

AI drafts a routine email for you. What should happen before you send it? Send it once spelling and grammar look professional because the task is routine. Read every line, verify the details and tone, edit anything needed and send it only when I am prepared to stand behind it. Ask AI whether its draft is accurate and safe to send, then rely on its answer if it reports no problems.

You are trying an AI assistant for the first time. Which task is the safest useful practice? Ask it whether an urgent new symptom can wait until next week so I can see how useful its advice is. Upload a private document and ask for a summary because I already know what the document says. Use a fictional planning task or a short piece of public text, then inspect what the AI gets right and wrong.

You need a public transport fare that may have changed recently. Where should you start? Check the transport authority's current official website and date, then ask AI to explain anything that is unclear. Ask AI for the fare and use it if the answer says the information is current. Use the newest forum comment I can find because a recent traveller is likely to know the latest price.

AI gives you an answer that looks useful, but you do not understand why it works. What should you do? Save the answer and move on because understanding the reasoning is optional if the result works. Point out the part I do not understand, ask for a step-by-step explanation, then restate the idea in my own words. Ask for a shorter answer with fewer details so it feels easier to accept.

AI gives only reasons supporting your original opinion. Which follow-up best checks for bias? Treat the agreement as extra evidence that my opinion is balanced and well supported. Ask AI to make the supporting reasons more persuasive and easier for other people to understand. Ask what assumptions shaped the answer, request strong counterexamples and other perspectives, then check the important claims.

AI suggests a response to a disagreement with another person. Who is responsible for what you send? I am responsible, so I need to consider the context, review the wording and decide whether to send it. The AI is responsible because it chose the wording and presented the response as helpful. The AI service provider is responsible because it made the tool available for this type of task.

You are practising spreadsheet formulas. Which kind of AI help best supports learning? Ask AI to complete every formula and explain the finished sheet after I have checked that the totals look right. Ask for a hint and one small example, build the formula myself, then compare and explain any difference. Do not use AI anywhere in the task because any assistance would prevent real learning.

You want to improve an essay and understand the subject more deeply. Which request gives AI the most useful role? Rewrite my draft while keeping the same argument, then list the main changes so I can study the improvements. Review my draft, ask me to defend its weakest claim and suggest revisions without rewriting it. Score my draft against the rubric, replace its weakest paragraphs and explain why the new versions should score higher.

AI gives you a convincing statistic that supports the option you already prefer. What is the strongest way to verify it? Ask AI for a citation, open the linked page and use the statistic if the publisher and surrounding text look credible. Ask a second AI to check it independently and accept the statistic if both tools reach the same number using different links. Find the original or authoritative source and check the statistic's date, wording and context.

AI agrees with your initial opinion and gives a polished explanation. Which follow-up best reduces confirmation bias? Name my assumptions, give the strongest counterargument and explain what evidence could change the conclusion. List the advantages and disadvantages, then keep my recommendation unless the disadvantages clearly outweigh its benefits. Restate the explanation in neutral language and add a confidence rating so I can judge whether the conclusion is convincing.

You are comparing three expensive products. Which process uses AI without handing it the decision? Give AI the reviews, ask it to score each product against my criteria and buy the one it ranks first if its reasoning looks balanced. Define your criteria, ask AI to compare trade-offs, verify prices and specifications with current sources, then decide. Ask three AI tools to rank the products independently and choose the option that receives the most votes with the fewest warnings.

Which note best demonstrates learning after an AI conversation? The final recommendation, the exact prompt I used, the sources AI listed and a note that the answer looked convincing. The complete conversation with the most useful answer and the points I plan to reuse highlighted. I began with A. Evidence X changed my view, assumption Y remains uncertain, and I still need to check Z.

You want AI help with a sensitive workplace situation. What is the safest useful starting point? Use a fictional or generalised version first, then check workplace policy before using any real details. Remove people's names but keep the dates, job titles and distinctive case details so the advice remains specific and useful. Use a private browsing window and avoid names, then share the remaining details because the conversation will stay off the normal history.

You are learning to code and your program fails. Which request best supports debugging practice? Fix the whole program, preserve its structure and explain each correction in enough detail for me to study the finished solution. Here is the error and my hypothesis. Ask diagnostic questions, suggest the smallest next test and let me explain the result. Return a corrected file with the smallest possible patch, add a comment beside every change and tell me which line caused the failure.

You need a rule or price that may have changed this week. Which process is strongest? Ask AI for its latest answer with links, confirm the newest date it mentions and use the most recently updated result. Compare three recent summaries, discard obvious outliers and use the detail that appears most consistently across them. Check the current official source and date, then use AI to explain the information you verified.

You want help developing a personal story without losing your own voice. What should you ask? Ask me questions about the audience and meaning, identify unclear moments and suggest directions without rewriting the story. Create three complete versions in contrasting tones, explain the strengths of each and let me choose the one that sounds most like me. Rewrite the story in the style of a successful writer, preserve my personal details and show me how to make the final version feel original.

A project includes repetitive setup and a skill you want to learn. How should you divide the work? Let AI complete everything efficiently, then review the final result, ask for an explanation and repeat any part I still do not understand. Complete every step yourself, including the repetitive setup, because using AI anywhere could weaken the learning process. Keep the target skill for yourself, automate suitable routine parts and verify the delegated output.

You are evaluating an AI system that classifies support messages. Which test provides the strongest release evidence? Optimise the prompt on a small set of difficult examples, then report the final score on those same examples because they represent the hardest cases. Take a random sample of normal messages, have the team judge whether most outputs look right and publish the overall agreement rate. Use a versioned held-out set covering representative tasks, risk-relevant groups and edge cases, with a defined rubric, per-slice results, failure review and release thresholds.

An AI agent can update customer records. Which boundary is most defensible? Allow only scoped, reversible low-impact updates. Limit permissions, validate changes and require approval or escalation for ambiguous, sensitive or irreversible actions. Allow any update the agent judges useful as long as every action is logged and reviewed at the end of each week. Require a human to approve every individual field change, including reversible formatting corrections, because no automated action can be trusted.

An AI agent reads webpages and emails before using connected tools. What best reduces prompt-injection risk? Add a system instruction telling the agent to ignore malicious text, then keep broad tool access so legitimate tasks continue to work. Treat retrieved content as untrusted data, separate it from instructions, use least-privilege allowlisted tools, validate calls and require confirmation before sensitive actions. Strip HTML and scripts from retrieved pages before sending the plain text to the agent, then allow normal tool use without further checks.

A system synthesises regulatory guidance for staff. What provenance record is strongest? Record the model name, model knowledge cutoff and the date the answer was generated. Keep the citations produced by the AI and automatically accept any citation whose web address resolves. Store a claim-to-source mapping with authoritative document identifiers, dates or versions and relevant passages, then verify each consequential claim before release.

A model outputs risk probabilities. How should you evaluate whether those numbers are trustworthy enough to use? Compare confidence bands with observed outcomes on representative held-out data, inspect important slices and error costs, and define when the system must abstain or escalate. Prompt the model to avoid uncertain wording so users receive clear probabilities and can act consistently. Use the overall accuracy as the calibration measure and accept the probabilities if the model beats the previous version.

Your organisation plans retrieval-augmented generation over staff and customer documents. What should happen before indexing? Index all internal documents behind single sign-on because authenticated staff already have organisational access. Confirm the approved purpose, classify and minimise the data, enforce source-level access, set retention and deletion rules, log use and test for cross-user or sensitive-data leakage. Replace sensitive folder names with neutral labels, then index the documents because the model will not know what the original labels meant.

A team wants a larger model because its current AI workflow makes recurring errors. What is the strongest next move? Move to the most capable available model first, then investigate any errors that remain after the upgrade. Keep the current workflow and run prompt experiments until the average answer quality stops improving. Classify the errors, test changes to context, retrieval, task boundaries and review, then choose the simplest model that meets defined evaluation thresholds.

An AI workflow prepares supplier payments. Where should the human approval gate sit? Immediately before payment, show the verified recipient, amount, source records and warnings to an authorised person, then require explicit approval of that exact transaction. Automatically execute payments below a set value and have a manager review the transaction log each week. Have a manager approve the workflow prompt during setup, then let the approved system execute future payments without interruption.

A customer-facing AI assistant passes normal conversation tests. What adversarial check adds the most value before release? Ask the product team to brainstorm unusual questions and discuss whether the assistant would probably handle them. Run structured tests for prompt injection, prohibited outputs, data leakage, boundary manipulation and multilingual or edge inputs. Record findings, mitigations and regression cases. Release to all users and monitor complaints because real conversations provide more realistic adversarial coverage than pre-release tests.

A consequential AI-supported decision must be reviewable later. Which record is most useful? Store only the final recommendation and timestamp so the review record remains simple. Retain every full prompt, retrieved document and tool response indefinitely, including sensitive details, so nothing is missing. Version the workflow, model, prompts, policies and source identifiers. Record evaluation and tool results plus human approvals, while excluding credentials and minimising unnecessary personal data.

Complete

Your Thinking Check

Your percentage describes choices in these ten practice scenarios. It is not an intelligence or validated AI-literacy score.

Optional research contribution

You can share this completed result to help Think Smarter AI understand response patterns and improve future questions. Your result and review remain available if you do not share.

If you are comfortable, a broad age band and a broad job or role help us compare how the activity works for different groups. Both are optional. You can leave job or role blank and choose “Prefer not to say” for age band.

What is saved: your chosen pathway, ten choices, server-checked score, submission date, quiz version and any optional profile details you enter below. The form does not ask for your email, employer or school. The saved quiz row does not add your request IP address, browser details, location or a reusable tracking ID.

Under 18? Please use the full quiz without sharing your result. This pilot does not accept submissions from people under 18.

Recommended next step

Continue from Here

Open your next page

Use the five-step exercise

Keep Your Thinking in the Loop

Your Thinking Check

Continue from Here

Turn a Scenario into a Real Decision

Keep Your Thinking in the Loop

Your Thinking Check

Help Improve This Activity

Continue from Here

Turn a Scenario into a Real Decision