News by Levels · Level 3 · For teachers

Teacher guide & answers

In this lesson: Assess what confidence scores do and do not establish.

Google Introduces Bounded AI Decisions

How to use these two pages

Give learners the student worksheet. It contains the story and questions, without the answers. Keep this guide for yourself. Use it to lead the activity and explain the answers afterwards.

Student worksheet

A simple 15–20-minute lesson

Before reading · 2 minutes
Show the title. Ask: “What do you think this story is about?”
Read and answer · 8–10 minutes
Give out the student worksheet. Learners read for the main idea, then reread to answer the 2 questions. They can compare answers with a partner.
Discuss and write · 5–8 minutes
Check the answers below together. Ask learners to point to the clue in the story, then finish with the open question.

Adjust the support: Let learners use word help or Level 1 if needed. For a challenge, compare the wording in another level. This activity is not a proficiency test.

Answers to the 2 questions

The question numbers match the student worksheet. Each answer includes a reason and the clue learners can find in the story.

What determines the answer options?
Correct answer: C — They are predefined
Why: The first paragraph describes evaluating predefined options.
Clue in the story: Google announced Gemini Decision 1 on November 11 for evaluating predefined options through its Decisions API. A single request can include up to 1,000 questions.
What is a supported use of an uncertain score?
Correct answer: B — Route an ambiguous case to a human
Why: The second paragraph gives human-review routing and explicitly rejects guaranteed correctness.
Clue in the story: The model returns probabilities and confidence scores instead of decoding an answer token by token. Those scores can help software route ambiguous cases to a human reviewer; they do not guarantee that every decision is correct.

The open question: accept different answers

Ask learners: Explain why a confidence score should not be reported as proof of correctness.

One possible answer: The model returns probabilities and confidence scores for predefined options. Software can use them to send ambiguous cases to a human reviewer, but the text explicitly says they do not guarantee that every decision is correct.

This is an example, not the only correct answer. Look for a relevant idea, a reason and the learner’s own words. Learners can answer aloud or write one or two sentences.

Grammar to discuss

Purpose rather than certainty

can help

Can help describes a possible useful role; it does not promise success in every case.