Testing AI Literacy with Google’s “Polar Peaks” AI Quest

Can you use AI to investigate a difficult scientific question, or can you only recognize familiar words such as “glacier,” “climate,” and “satellite”? Google’s “Polar Peaks” AI Quest offers a useful practical test: start with an unfamiliar polar-science problem, ask an AI system to help, inspect its reasoning, verify important claims, and record what remains uncertain.

This article gives you a repeatable way to complete that exercise from zero. It does not report a completed quest, a score, or a result from a particular account. Instead, it defines what to do, what to save, what to check, and how to distinguish a useful answer from a fluent but unreliable one.

What this test is actually measuring

AI literacy is not the ability to name machine-learning concepts. In this exercise, it means being able to use an AI system deliberately while understanding its limits.

A successful attempt should show that you can:

• translate a broad scientific topic into answerable questions;
• give the AI enough context to produce a useful first explanation;
• separate observations, interpretations, and guesses;
• ask for evidence instead of accepting confident wording;
• notice missing definitions, hidden assumptions, and contradictions;
• revise your question after learning something new; and
• state what you still do not know.

The result is therefore a process record, not just a final paragraph. A polished answer with no verification is weak evidence of AI literacy. A partly incomplete answer with clearly documented checks can be strong evidence.

Before you begin: define the evidence boundary

Keep this distinction visible throughout the quest:

This boundary prevents a common failure: copying the AI’s wording and accidentally upgrading a possibility into a fact.

Materials and setup

You need:

• access to the Google “Polar Peaks” AI Quest;
• an AI chat tool that permits you to enter and export text;
• a notes file or spreadsheet for recording prompts and answers;
• a browser or reference tool for checking claims; and
• about 25 minutes without entering personal, confidential, or account information.

The exact quest interface, wording, and available model may change. Record the date, the tool name, and the quest instructions you saw. Do not treat this article as a substitute for the instructions shown in your own session.

Step 1: capture the starting question

Copy the quest’s initial challenge into your notes. If the interface does not provide a single question, write a one-sentence version in your own words.

Use this template:

Date:
Quest name: Polar Peaks
AI tool and model, if shown:
Starting challenge:
What I already know:
What I do not know:
What a good answer must explain:

Do not search for an answer yet. The purpose of this first record is to show your starting point. It also makes later improvement measurable.

Step 2: decompose the scientific problem

Complex topics become manageable when you split them into smaller questions. Ask the AI to help with structure, but do not ask it to decide what is true.

Use a first prompt like this:

You are helping me investigate the Polar Peaks challenge as a beginner.

First, decompose the challenge into:
1. key terms I must define;
2. observations or measurements involved;
3. possible explanations;
4. information needed to distinguish those explanations; and
5. questions that cannot be answered from the current context.

Do not present guesses as facts. Mark each item as definition, observation,
hypothesis, evidence needed, or unknown. Ask me no more than three
clarifying questions at the end.

Read the response line by line. Highlight any statement that sounds factual but has no source, date, measurement, or definition. Those highlights become your first investigation list.

Step 3: build a small evidence table

Convert the AI’s response into a table. This is slower than copying a summary, but it exposes the difference between information and interpretation.

Use “No” honestly. A blank verification column is more useful than a fabricated citation.

Step 4: ask for competing explanations

A good scientific investigation does not search only for evidence supporting the first idea. Ask the AI to generate alternatives and weaknesses in its own explanation.

For each proposed explanation, provide:

- what it would predict;
- what observation would support it;
- what observation would weaken it;
- one plausible alternative explanation;
- what information is missing.

Use cautious language. If the information is not available, write
“not established” instead of inventing a value or source.

Now inspect whether the alternatives are genuinely different. “More snow” and “different snow conditions” may be related rather than independent explanations. If you cannot explain how two hypotheses would produce different observations, your question is not yet testable.

Step 5: verify the highest-risk claims

Do not try to verify every sentence equally. Prioritize claims that could change your conclusion:

• numbers, dates, rankings, and trends;
• claims about causation;
• named studies, datasets, organizations, or locations;
• technical definitions that affect interpretation;
• statements that use words such as “always,” “proves,” or “the main reason.”

For each priority claim, ask the AI to expose its basis:

List the five claims in your previous answer that are most important
to the conclusion. For each one, state:

- the exact claim;
- whether it is a fact, interpretation, or hypothesis;
- the source or data required to verify it;
- what would make the claim unreliable.

Do not create links, citations, or study names if you do not know them.

Then check the claims independently using the sources available to you. A source is not automatically strong because the AI names it. Open it, confirm that it exists, and check whether it supports the precise claim being made.

Step 6: test for hallucination and overconfidence

A hallucination is an AI-generated statement that appears informative but is unsupported, false, or mismatched to the question. It may be a fabricated source, an incorrect number, a real source used to support a different claim, or a plausible explanation presented as established fact.

Run a controlled challenge:

Review your answer as a skeptical editor.

Return four sections:
1. claims directly supported by information in this conversation;
2. claims that require external verification;
3. claims that may be too confident or causally stronger than the evidence;
4. details you may have invented or cannot substantiate.

For section 4, be explicit. Do not replace missing evidence with a guess.

Compare this response with your evidence table. If the AI now retracts or weakens a claim, record the change. The change is not a failure of the exercise; failing to notice the change is the failure.

Step 7: make the AI explain uncertainty

Uncertainty is not the same as ignorance. It can arise because measurements are noisy, definitions differ, the sample is limited, or several explanations fit the available observations.

Ask:

Rewrite the current explanation for a careful beginner.

Separate:
- what is directly observed;
- what is inferred;
- what is plausible but unconfirmed;
- what cannot be concluded;
- what additional measurement or source would reduce uncertainty.

Avoid percentages unless they are calculated from identified data.

Look for concrete uncertainty, not decorative phrases such as “it depends.” A useful answer identifies what it depends on and what information would resolve the dependency.

Step 8: produce a short evidence-bounded answer

Write the final response in your own words. You may ask the AI to edit it, but keep the evidence labels.

Using only the verified items in my table, draft a 180-word answer.

Requirements:
- define unfamiliar terms;
- distinguish observations from explanations;
- include no unsupported numbers or citations;
- name the strongest remaining uncertainty;
- state one next investigation;
- label any unresolved explanation as a hypothesis.

Before accepting the draft, compare every factual sentence with your notes. Remove any sentence that cannot be traced to evidence you actually checked.

Verification checklist

You have completed the practical test when you can answer “yes,” “no,” or “unknown” to each item:

• Did I preserve the original challenge and my starting assumptions?
• Did I split the challenge into definitions, observations, hypotheses, and unknowns?
• Did I ask for at least one competing explanation?
• Did I identify the claims that matter most to the conclusion?
• Did I independently inspect the evidence for those claims?
• Did I record units, dates, scope, and limitations where relevant?
• Did I test whether the AI invented or overstated anything?
• Did I rewrite the result using only verified or clearly labeled material?
• Can I state what evidence would change my current conclusion?

If you answered “unknown” to an item, keep it as an explicit limitation. Do not convert it into “yes” because the final prose sounds convincing.

Common failure cases

Recognizing vocabulary instead of investigating

You may know that glaciers, ice sheets, snow, albedo, and satellites are relevant, yet still be unable to explain how a particular observation was measured or what would disprove a proposed explanation. Fix this by requiring every important term to connect to a question, measurement, or decision.

Accepting a fluent summary

Fluency is a presentation property, not evidence. Ask the AI to mark the boundary between information supplied in the conversation and information it would need to verify externally.

Confusing correlation with causation

If two variables change together, that does not establish that one caused the other. Add alternative causes and ask what observation would separate them.

Using a source without checking its scope

A source about one region, period, or measurement method may not support a claim about all polar environments. Record the source’s scope before using it.

Letting the AI invent missing data

Requests for “a realistic estimate” or “a likely value” can produce numbers that look scientific but are not measurements. If a number is not required, omit it. If it is required, identify the dataset and calculation.

Changing the question without noticing

During a conversation, a broad question can silently become a narrower one. Keep the current question in your notes and compare it with the original challenge before writing the conclusion.

What this test cannot prove

One completed quest cannot measure your general scientific knowledge, research ability, or performance across different AI systems. It also cannot establish that every answer produced during the session was correct.

The exercise is a small behavioral sample. It tests whether you can apply a disciplined workflow to one complex topic. A stronger evaluation repeats the process with a different subject, a different model, and a deliberately ambiguous source set.

Do not use the exercise as a substitute for expert review in medical, legal, financial, safety-critical, or research-publication decisions. In those settings, verification requirements are higher than this beginner workflow.

Final deliverable: your mistake log

Finish by saving a short list of mistakes to investigate further. Include the original wording, why it was risky, how you checked it, and what you will do differently next time.

Mistake:
Original AI statement:
Why I initially accepted it:
What check exposed the problem:
Corrected wording:
Follow-up question:
New habit for the next quest:

The most useful outcome of Testing AI Literacy with Google’s “Polar Peaks” AI Quest is not a perfect answer. It is a visible record showing how you moved from an AI-generated possibility to a bounded conclusion—and where the evidence still runs out.

Continue with the Agent Lab Journal guides or look up unfamiliar terms in the glossary.