Comprehension test | Hands-on AI-Assisted Development: Mechanical and Electrical track, Session 1

15 questions, each with 4 choices. You may refer to the materials while you answer.

When you finish, click "Compile answers" at the bottom of the page. Your answers appear as text. Submit them as the instructor explains. We will tell you how to submit on the day.

The questions contain no client names or real figures. Do not write any in your answers either.

Which description is closest to how generative AI writes text?

It searches an internal document database and returns the best-matching text as is. It converts the meaning of the question into a logical formula, calculates the correct answer and returns it. It extends the text one word at a time, choosing a likely word to follow the text so far. It picks a matching answer template prepared in advance and fills it in.

Which task is generative AI most likely to help with?

Having it draft several possible openings for a report that is hard to start from a blank page. Having it judge whether this task violates the latest version of the internal rules. Having it calculate the defect rate from measured values and putting that figure directly into a document you submit. Having it summarize past exchanges with a client and finalize them as agreements for the next meeting.

Which of the following is generative AI bad at?

Rewriting Japanese text in English. Turning bullet points into sentences. Summarizing a long text into a short one. Correctly saying "I don't know" about things it does not know.

Which is the best description of hallucination?

Processing stops in the middle of an answer, and the text is cut off. The AI presents untrue content in a way that makes it look like fact. The AI answers the same question in different wording each time. The question is so long that its first half is ignored.

In which situation must you not use a generative AI answer as is?

When it points out typos in a text you wrote. When it rewrites the headings of meeting notes. When you ask its opinion on whether your explanation is hard to follow. When you put it in a document for a client as the basis for judging conformity to a standard.

When you do not get the answer you expected, what is a good thing to try first?

Ask again and add who the text is for, about how long it should be, and what it will be used for. Send the same question again and again until a good answer comes back. Translate the question into another language before you send it. Rewrite part of the answer yourself and send the same question again.

Which is the right way to think about responsibility for content that generative AI produces?

Because the AI produced it, the user is not responsible even if it is wrong. If sources are shown, you may use the content without checking it. Whoever uses the content is responsible for what goes out as a submission or deliverable. For internal documents, you do not need to check the content.

When you consult AI about your own work, how should you write the client name?

Write the official name as is. Shorten it to an initial, such as "Company A". Generalize it to the level of industry and process, such as "the design process in the automotive industry". Leave out the company name and write the contact person's name instead.

Which statement about the data you give AI to read in training or in trials is correct?

Use real work files with only the file names changed. Do not use real data, real files or screenshots of real screens. Real data is fine if you shift the numbers slightly. Anything in an internal shared folder is fine to use.

Q10

Which is an appropriate way to write a message that asks AI about a work problem?

Copy specific values such as dimensional tolerances and settings exactly and give them to the AI. Avoid specific values, and write at the level of the goal, where you are stuck and what result you want. Paste the relevant page of the specification as is. Add the drawing number so that it is clear which project it is about.

Q11

What must you always do before you submit or share a text you have written?

Get your manager's approval, then check the content. Have the AI read it again and judge whether there is a problem. Check that the length is within the limit. Check by yourself that the content is safe to go outside the company.

Q12

You work in a client's environment. Which point needs special attention?

You must follow the rules of the client site where you work, not only your own company's rules. If your company allows a tool, you can use it freely at the client site too. The client site's rules never take priority over your own company's rules. If you personally think it is fine, you do not need to check.

Q13

You notice that you entered information that should have been kept hidden. What should you do?

Nothing. Deleting the conversation is enough. Send another question that overwrites the same content, to dilute the record. Report it promptly, following your company's procedure. Decide to be careful next time and just make a note of it.

Q14

Which statement about the basis for using GitHub Copilot at work is correct?

If you sign up personally, you can use it freely for work at the client site. Whether you may use it, and how far, follows the agreements between your company and the client site. If you stay within the free tier, you do not need to report it. Permission is needed only when you use it from outside the company network.

Q15

You start using what you learned today from tomorrow. Which approach is most appropriate?

Start with tasks that have little impact, and check the output yourself. Start with important deliverables to show the benefit. Expand the scope on your own judgment without consulting anyone. Use it first for tasks that need especially high accuracy.