Learning path · Level 3 · Tutorial 3

Checking AI Analysis of Text and Tables

Analyze short text and small tables with an assistant, then check whether each conclusion is supported and whether the calculations are correct using a calculator or simple manual checks.

  • Level: Intermediate
  • Duration: 20 min
  • 7 steps
Start tutorial Back to the path
Checking AI Analysis of Text and Tables

Step 1 of 7

Ask for an Analysis You Can Check

Create a request that makes an assistant’s analysis of Cedar Learning Center evidence easier to inspect.

Start with a bounded source

Cedar Learning Center: source notes

Use this fixed case material for this tutorial. Treat it as the complete source: do not add facts from outside it.

P1: Registrations counts registered people within each session. Attended counts those registrants who attended. A person may appear in multiple sessions.

P2: Reminder emails began in May. The venue also changed in May. Organizers believe reminders improved attendance.

P3: Feedback was optional. Feedback response counts are unavailable. The meaning of the Positive column is unspecified.

T1

T1 records four Cedar Learning Center sessions. An em dash means the value is unavailable; it is not a recorded zero.

Cedar Learning Center session records
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Design an inspectable response

Four categories, four jobs

Ask the assistant to keep four kinds of statements separate.

Evidence is what P1–P3 explicitly state or what T1 records. Interpretation is what an analyst thinks those materials suggest. Assumptions are premises used without support in the supplied material. Unresolved questions identify information still needed.

Separation makes a response easier to inspect. It does not make every statement true: the evidence and references still need review.

Keep statement types apart

A response can be organized as a visible path from source material to four different kinds of statements.

A source-note stack and a table connect by arrows to four distinct panels representing evidence, interpretation, assumptions, and unresolved questions.

Use this order: Evidence, Interpretation, Assumptions, Unresolved questions.

Tip

Make evidence traceable

Require a source reference with each evidence statement. Use paragraph references such as P2 and table references such as T1:R3 (Attended). A reference helps a reviewer find the material; it is not proof by itself that the claim is accurate.

See the format in use

A request pattern

A useful request gives the assistant a task, boundaries, and a reviewable output format. This pattern tells it not to turn an interpretation or an assumption into an asserted source fact.

Example

Example request

Analyze the Cedar Learning Center source notes P1–P3 and table T1 using only the supplied material. Do not use outside information.

Organize the response under these headings, in this order:
Evidence
Interpretation
Assumptions
Unresolved questions

Under Evidence, state only what the passages explicitly say or T1 records, and attach a reference such as P2 or T1:R3 (Attended) to each point. Label interpretations as interpretations. Identify premises that are not established as assumptions. State what information is still needed as unresolved questions. Do not treat the headings themselves as proof that a claim is correct.

Example

Example assistant response

Evidence

  • T1 records 15 attended in April Morning and 20 attended in May Morning. (T1:R1 (Attended); T1:R3 (Attended))

Interpretation

  • The Morning session attendance records appear higher in May than in April.

Assumptions

  • Calling this difference an overall attendance improvement assumes that these two Morning sessions represent all relevant attendance records.

Unresolved questions

  • What does the Positive column represent? P3 says its meaning is unspecified.

The headings reveal what kind of statement each line is. They do not remove the need to check the cited passages and cells.

Draft your analysis instructions

Your turn: write a checkable request

Draft a short request to the assistant about the Cedar Learning Center material. Require an analysis that separates source facts from interpretations, assumptions, and questions that remain open. Include the four prescribed output headings in their specified order, require source references for evidence, and make clear that organization is not proof.

Do not analyze the case itself; write the request you would give the assistant.

Write at least 180 characters (0/180).

Step 2 of 7

Trace Each Claim to Its Source

Locate the exact passage or table cell behind a claim, then decide whether the source supports, contradicts, or does not resolve the claim.

Read the sources

Cedar Learning Center source material

Use these fixed source notes and table as the material to check. A source reference tells a reader exactly where a claim can be inspected.

Source notes

P1: Registrations counts registered people within each session. Attended counts those registrants who attended. A person may appear in multiple sessions.

P2: Reminder emails began in May. The venue also changed in May. Organizers believe reminders improved attendance.

P3: Feedback was optional. Feedback response counts are unavailable. The meaning of the Positive column is unspecified.

T1

T1 records four Cedar Learning Center sessions. A dash (—) is an unavailable value.

Session records for April and May
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Build a claim-to-source map

Split, then trace

Check one claim at a time. A sentence can contain several claims, and each may need a different reference.

For a text claim, locate the exact paragraph. For a number, locate the row and column that contain it. Then compare the full wording of the claim with the source. A reference that mentions the topic is not enough: it must support the claim's scope and detail.

Example

Worked tracing example

Statement: “May Morning had 25 registrations, and organizers believe reminders improved attendance.”

Split it into two checks:

  • “May Morning had 25 registrations.” → T1:R3 (Registrations)
  • “Organizers believe reminders improved attendance.” → P2

The second reference supports what the organizers believe. Keep that wording; do not replace it with a stronger claim about what the records prove.

Map each claim separately

A claim-to-source map makes separate support paths visible.

Two claim cards connect separately to a paragraph and to one highlighted cell in a small table.

One claim may point to a paragraph, while another points to a specific table cell.

Judge what the source actually says

Choose a precise verdict

After locating a reference, assign a verdict. The verdict concerns the claim as written—not whether its topic appears somewhere in the material.

Verdict guide

Use the source wording and values to decide whether a claim is supported, contradicted, or still unresolved.

A reference must support the whole claim.
VerdictUse whenExample check
SupportedThe source states the whole claim.P1 states that a person may appear in multiple sessions.
ContradictedThe source conflicts with the claim.T1:R4 (Attended) records 15, not a claim of 20.
UnresolvedThe source does not establish a needed detail.P2 says both events occurred in May, not which happened first.

Practice tracing claims

Claim check

Match each claim to its source reference and verdict.

Tap an item and then its matching pair.

Step 3 of 7

Find Gaps Before Drawing Conclusions

Identify missing values, unclear labels, and counting-unit limits that make a conclusion unreliable.

Spot the gaps in T1

Inspect the limits in the source

Before accepting a conclusion, inspect what the source does not establish.

P1: Registrations counts registered people within each session. Attended counts those registrants who attended. One person may appear in more than one session.

P3: Feedback was optional, feedback response counts are unavailable, and the meaning of Positive is unspecified.

These limits matter even when a table looks complete. A blank-looking unavailable value, an unclear category, and repeated people across sessions each require a different clarification.

T1

T1 records four Cedar Learning Center sessions. Keep the heading Positive exactly as supplied: P3 does not define it.

Cedar Learning Center session records
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Read symbols and units carefully

A comparison of an unavailable dash and a recorded zero, plus one person linked to two session records.

A dash means the value is unavailable; it is not evidence of a recorded zero. Because one person can appear in multiple sessions, session records cannot automatically be counted as distinct people.

Convert gaps into unresolved questions

Ask for the clarification that is missing

Use a precise question for each gap.

  • Unavailable value: T1:R2 (Positive) is —. Do not replace it with 0, and do not infer why it is unavailable.
  • Unclear category: P3 leaves Positive undefined. A number under that heading cannot safely be called a satisfaction count, a favorable rating, or any other convenient category.
  • Missing denominator: A feedback percentage needs the number of feedback responses in the same scope. P3 says those counts are unavailable.
  • Counting unit: P1 permits the same person in more than one session. Registration and attendance records across sessions are not automatically counts of unique people.

Ask for the missing definition, count, or identifier. Do not fill the gap with an assumption.

Example

Example: record the limit, not a guess

Evidence

  • P3 says feedback was optional, response counts are unavailable, and the meaning of Positive is unspecified.
  • P1 says a person may appear in multiple sessions.

Interpretation

  • The supplied material does not establish a feedback percentage or a unique-person total across sessions.

Assumptions

  • Calling Positive “satisfied participants” would assign a meaning not supplied by the source.
  • Treating every session record as a different person would assume no repeat participants.

Unresolved questions

  • What does Positive measure?
  • How many feedback responses were received for each comparison scope?
  • Which records belong to the same person across sessions?

Check a conclusion before accepting it

Choose the necessary clarification

An assistant writes: “May had a higher feedback percentage than April, and all registration records in T1 represent distinct people.” Which clarification is needed before this conclusion can be checked?

Step 4 of 7

Recalculate Counts and Totals

Independently rebuild Cedar Learning Center totals from T1, reconcile monthly and overall values, and preserve the difference between attendance records and unique people.

Read the records before calculating

Start with the source and its unit

To check a total, return to the source values—not to the assistant’s intermediate arithmetic.

P1 states that Registrations counts registered people within each session, and Attended counts those registrants who attended. A person may appear in multiple sessions. Therefore, a sum of Attended values is a count of attendance records across sessions, not a count of unique people.

T1

Use T1 as the original record for every calculation in this step.

Cedar Learning Center session records
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Keep categories separate when totaling

A four-record diagram groups two sessions in April and two in May, with separate paths for registration values and attendance values.

Group the relevant session rows by month, then add values from one column at a time.

Rebuild, then reconcile

A repeatable total check

Use the same checking sequence each time:

  1. State the scope: which rows qualify?
  2. Identify each qualifying row exactly once.
  3. Take values from only the requested column.
  4. Add the source values yourself.
  5. Check that monthly results reconcile with an overall total.

As a basic consistency check in T1, an Attended value cannot be greater than Registrations for the same row.

Example

Worked check

Claim to check: April registrations total 50.

Trace the values to T1:R1–R2 (Registrations). Rebuild the arithmetic from those cells:

20 + 30 = 50 registrations

This supports the claim for April registrations. It does not say anything about attendance records, because those are in a different column.

Tip

Use an independent calculation

For an independent check, enter the source values from T1 into a calculator yourself. Asking the assistant to repeat its subtotal only repeats a result; it does not independently test the arithmetic.

Calculate from T1

Check the scope

T1 contains ___ listed sessions.

Check April

April attendance records total ___.

Check May

May attendance records total ___.

Check registrations

Across all listed sessions, registrations total ___.

Check attendance

Across all listed sessions, attendance records total ___.

Report the checked totals accurately

Final calculation audit

Write one concise audit result using T1 and P1. Include the number of listed sessions, the April and May attendance totals, the overall registrations total, and the overall attendance total. State why the attendance total must not be described as a count of unique people.

Write at least 110 characters (0/110).

Step 5 of 7

Check What Each Percentage Is Of

Define the percentage before using a calculator

Name the part and the whole

A percentage answers a part-of-a-whole question. Before calculating, name the scope: which sessions are included, what is the part, and what is the matching whole.

For an attendance rate, use Attended / Registrations × 100 for the same stated sessions. P1 defines Registrations as people registered within each session and Attended as those registrants who attended. When sessions are combined, describe the numerator as attendance records, not unique people: one person may appear in more than one session.

A number can be arithmetically correct yet answer the wrong question if its denominator comes from a different scope.

Part-to-whole check

A group of thirty highlighted person icons inside a larger group of fifty person icons.

For April, 30 attendance records are the part and 50 registrations are the matching whole.

Use totals, not a simple average

Source table: T1

T1 contains the source values. Keep each column separate: Registrations is the denominator for an attendance rate; Attended is the numerator. The Positive column is not an attendance count.

Cedar Learning Center sessions in April and May.
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Example

Combine counts before dividing

First calculate each session using its own counts:

15 / 20 × 100 = 75%

15 / 30 × 100 = 50%

These session rates should not simply be averaged when the session sizes differ. Instead, add the relevant attendance records and registrations first, then divide:

(15 + 15) / (20 + 30) × 100 = 60%

The simple average of 75% and 50% is 62.5%, which does not represent the combined April attendance rate.

Check inputs and limits

Diagnose the denominator error

An assistant writes: “For April, the attendance rate is 30 / 30 × 100 = 100%.” What is the error?

Select matching source values

Which source values should be combined to calculate the May attendance rate?

Do not invent a feedback percentage

P3 says feedback was optional, feedback response counts are unavailable, and the meaning of Positive is unspecified. What verdict applies to the claim “The feedback percentage was 58%”?

Independently verify the three rates

Calculate April

Using T1:R1–R2, calculate the April attendance rate:

(15 + 15) / (20 + 30) × 100 = ___ %

Calculate May

Using T1:R3–R4, calculate the May attendance rate:

(20 + 15) / (25 + 25) × 100 = ___ %

Calculate the combined rate

Using all rows in T1, calculate the combined attendance rate:

(15 + 15 + 20 + 15) / (20 + 30 + 25 + 25) × 100 = ___ %

Step 6 of 7

Keep Conclusions Within the Evidence

Revise a conclusion so it keeps checked findings while avoiding an unproven explanation.

Separate the observation from the explanation

Observation is not explanation

A checked difference tells you what the records show, not automatically why it happened.

For Cedar Learning Center, the April attendance rate is 30 / 50 × 100 = 60%, using T1:R1–R2 (Registrations, Attended). The May attendance rate is 35 / 50 × 100 = 70%, using T1:R3–R4 (Registrations, Attended). The records therefore show a 10-percentage-point increase.

P2 says reminder emails began in May, the venue also changed in May, and organizers believe reminders improved attendance. That is evidence of the organizers’ belief and of both changes occurring in May. It is not proof that reminders caused the increase.

Same-month changes do not identify a cause

A timeline compares April attendance records with higher May attendance records. An email reminder icon and a venue-change location icon both appear in May, with no causal arrow.

T1 shows a higher attendance rate in May. P2 places both reminder emails and a venue change in May, so this comparison does not isolate the effect of reminders.

Build a bounded conclusion

Example

Model: keep each layer visible

Evidence
T1:R1–R2 (Registrations, Attended) gives an April attendance rate of 30 / 50 × 100 = 60%. T1:R3–R4 (Registrations, Attended) gives a May attendance rate of 35 / 50 × 100 = 70%. P2 reports that reminder emails began in May, the venue changed in May, and organizers believe reminders improved attendance.

Interpretation
The overall attendance rate was 10 percentage points higher in May than in April.

Assumptions
Saying that reminders caused the increase assumes that the venue change and other possible differences did not affect attendance.

Unresolved questions
The supplied material does not show whether reminders, the venue change, or another factor explains the difference.

Tip

Be specific about uncertainty

A precise revision does not need a vague word such as “perhaps.” State the exact limit instead: the May comparison cannot determine the effect of reminders because the venue also changed in May.

Check the wording

Choose the defensible conclusion

Which conclusion stays within the available evidence?

Apply the evidence limit

Rewrite the overclaim

Rewrite this conclusion in 2–4 sentences:

“Reminder emails caused Cedar Learning Center’s May attendance rate to rise.”

Use the checked April and May rates from the previous screen. Distinguish the organizers’ belief in P2 from a verified cause, and name the venue change in May as the specific reason the explanation remains unresolved.

Write at least 120 characters (0/120).

Step 7 of 7

Complete a Small Analysis Audit

Audit packet: Cedar Learning Center

Audit claims one at a time

Use one audit path for each claim: locate its source, check any arithmetic independently, check whether a label or counting unit is defined, and then decide how far the conclusion can go. A polished response can contain both reliable and unreliable claims.

A visual flow from source notes and a table through claim-by-claim checks to a limited conclusion.

Do not accept or reject a whole assistant response at once; audit each claim separately.

Source notes

P1: Registrations counts registered people within each session. Attended counts those registrants who attended. A person may appear in multiple sessions.

P2: Reminder emails began in May, the venue also changed in May, and organizers believe reminders improved attendance.

P3: Feedback was optional, feedback response counts are unavailable, and the meaning of the Positive column is unspecified.

T1

Use the table name and row-and-column references when you trace numerical claims.

Cedar Learning Center session records
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Example

Fresh assistant-style response

Evidence

  • May has 35 attendance records, because T1:R3–R4 (Attended) are 20 and 15.
  • The May attendance rate is 35 / 90 × 100 = 38.9%.

Interpretation

  • Reminder emails caused attendance to improve in May.

Assumptions

  • Positive means satisfied participants, so 29 / 50 × 100 = 58% of registrants gave positive feedback.

Unresolved questions

  • There are no unresolved questions: May served 35 unique people.

Two quick audit checks

Keep the valid claim

Which claim from the response can be retained as Supported?

Correct the denominator

Using a calculator or manual work, complete the independently checked May combined attendance rate:

35 / 50 × 100 = ___ %

Final application

Source notes for your audit

P1: Registrations counts registered people within each session. Attended counts those registrants who attended. A person may appear in multiple sessions.

P2: Reminder emails began in May, the venue also changed in May, and organizers believe reminders improved attendance.

P3: Feedback was optional, feedback response counts are unavailable, and the meaning of the Positive column is unspecified.

T1

Use T1 as the source for your traceable calculation.

Cedar Learning Center session records
RowMonthSessionRegistrationsAttendedPositive
R1AprilMorning20158
R2AprilEvening3015—
R3MayMorning252012
R4MayEvening25159

Example

Response to audit

Evidence

  • May has 35 attendance records, because T1:R3–R4 (Attended) are 20 and 15.
  • The May attendance rate is 35 / 90 × 100 = 38.9%.

Interpretation

  • Reminder emails caused attendance to improve in May.

Assumptions

  • Positive means satisfied participants, so 29 / 50 × 100 = 58% of registrants gave positive feedback.

Unresolved questions

  • There are no unresolved questions: May served 35 unique people.

Complete the small analysis audit

Write a corrected, traceable audit of the response above. Use these headings in this exact order: Evidence, Interpretation, Assumptions, Unresolved questions.

Your audit must:

  • use Supported, Contradicted, or Unresolved for each claim you assess;
  • retain the valid May attendance-record finding with an exact source reference;
  • independently recalculate the May attendance rate with its numerator, denominator, and result;
  • identify the unsupported premises in the feedback claim or unique-people claim;
  • replace the causal reminder claim with a bounded conclusion that accounts for P2;
  • state at least one specific unresolved question.

Use the source packet and your own arithmetic, not a repeated assistant calculation.

Write at least 280 characters (0/280).

Compare and close

Example

Model audit: verdict every claim separately

Evidence

  • Supported — “May has 35 attendance records.” T1:R3–R4 (Attended) records 20 and 15: 20 + 15 = 35 attendance records.
  • Contradicted — “The May attendance rate is 35 / 90 × 100 = 38.9%.” The denominator does not match the May scope. T1:R3–R4 (Registrations) records 25 and 25: 25 + 25 = 50. The independently checked May rate is 35 / 50 × 100 = 70%.
  • Supported — The April attendance rate is 30 / 50 × 100 = 60%, using T1:R1–R2 (Registrations, Attended). Supported — The May rate is 35 / 50 × 100 = 70%, using T1:R3–R4 (Registrations, Attended).

Interpretation

  • Unresolved — “Reminder emails caused attendance to improve in May.” The records show a 10-percentage-point difference between the checked April and May rates, but P2 also says the venue changed in May. P2 records the organizers’ belief; it does not establish the cause.

Assumptions

  • Unresolved — “Positive means satisfied participants, so 29 / 50 × 100 = 58% of registrants gave positive feedback.” P3 leaves Positive undefined and says feedback response counts are unavailable. This claim assumes both a meaning for Positive and a matching feedback denominator.
  • Unresolved — “May served 35 unique people.” The 35 is an attendance-record total from T1:R3–R4 (Attended). P1 says a person may appear in multiple sessions, so it does not establish a unique-person count.

Unresolved questions

  • Contradicted — “There are no unresolved questions.” P1 and P3 leave necessary information unavailable.
  • What does Positive measure? How many feedback responses were received in each relevant scope? Did the same people attend more than one session? What evidence could separate a reminder effect from the venue change?

Summary

Audit checklist

Use this compact checklist whenever you inspect an assistant analysis of supplied material.

  • Split the response into separate claims; do not grade it as one all-or-nothing answer.
  • Trace each retained statement to an exact passage or table reference.
  • Recalculate totals and percentages from source values, with a matching numerator, denominator, scope, and unit.
  • Give each assessed claim a clear verdict: Supported, Contradicted, or Unresolved.
  • Name unsupported premises behind claims about undefined labels, unavailable values, or unique people.
  • State the verified observation, then name the specific limit on any explanation.

Step complete

Congratulations! You completed: Checking AI Analysis of Text and Tables

You completed a small analysis audit: preserve supported findings, correct arithmetic, identify unsupported premises, and keep conclusions within the available evidence.

Download the App now to have access to + 5000 free courses, exercises, certificates and lots of content without paying anything!

  • 100% free online courses from start to finish

    Thousands of online courses in video, ebooks and audiobooks.

  • More than 60 thousand free exercises

    To test your knowledge during online courses

  • Valid free Digital Certificate with QR Code

    Generated directly from your cell phone's photo gallery and sent to your email

Cursa app on the ebook screen, the video course screen and the course exercises screen, plus the course completion certificate