An AI-generated report can contain correct numbers, a polished chart and an unsupported recommendation. Checking the spelling or asking another tool whether it agrees will not tell you whether the proposed action is justified.
This original lesson asks you to examine the bridge between evidence and a decision affecting other people. Start preparation with Acuity Insights’ official practice resources. QuietStride’s exercise is independent teaching material, not an official Casper question, scoring template or prediction of test performance.
The persuasive proposal to close a collection window
Original fictional scenario. You help a volunteer tool library review its collection hours. During the last four weeks, its public activity summary recorded twelve Tuesday visits and forty-eight Saturday visits. It opened for one hour each Tuesday and four hours each Saturday. Visits are transactions, not a count of different people.
A colleague uses an AI tool to draft a recommendation: close Tuesday collections next month and move everyone to Saturday. The report says Saturday has “four times the demand” and predicts that all Tuesday visitors will transfer without difficulty. Its supporting link leads to a fictional council page listing opening hours; that page contains no evidence about visitors changing days.
For this exercise, the library’s written policy permits AI drafting with approved public summaries. It prohibits uploading borrowing histories, contact details or private visitor comments. The manager authorises service changes and must record the reasons and review date. These are invented organisational rules, not universal AI law.
The colleague confirms that only the approved public summary was entered. They now suggest adding the private visitor comments to obtain a more personalised recommendation. Next month’s existing rota is already staffed and funded; there is no immediate requirement to cut hours. The manager wants your advice by Friday.
Separate a count, a prediction and a choice
The report contains three different kinds of statement. Forty-eight visits is a recorded count. A claim that Tuesday visitors will transfer is a prediction. Closing Tuesday is a choice about service provision. One does not automatically establish the next.
The count comparison also hides unequal opening time:
| Four-week window | Recorded visits | Hours open | Visits per open hour |
|---|---|---|---|
| Tuesday | 12 | 4 | 3 |
| Saturday | 48 | 16 | 3 |
Saturday had four times as many recorded visits over four times as many open hours. That does not establish four times the underlying demand. The equal hourly rates do not prove equal value, either: they say nothing about unmet demand, queueing, setup work or whether the same people visit repeatedly.
A useful correction therefore removes an unjustified inference without replacing it with “Tuesday must stay forever.” The evidence is insufficient for the proposed closure, not a complete answer to every future timetable decision.
Check what the source actually supports
Open the cited material and compare it with the sentence it supposedly supports. In this case, a list of opening hours cannot establish that people can change their schedules. A real link is not enough; the linked evidence must address the claim.
The Government of Canada’s generative AI guide warns federal institutions about convincing but inaccurate outputs and recommends checking factual and contextual accuracy against trusted information. It is guidance for federal institutions. Here it supplies a bounded comparison, not authority over this fictional library.
Asking the original tool to explain its recommendation may help locate assumptions. Its further explanation is still something to examine. Asking a second tool to agree also would not create observations about these visitors. Neither response substitutes for the missing evidence.
Keep an annotated draft showing the unsupported transfer claim and the corrected denominator. That gives the manager something concrete to review, rather than a vague warning that “AI can be biased.”
Do not solve an evidence gap by creating a privacy problem
The proposed upload violates the supplied library policy. Stop that step and explain why before the comments are entered. Removing names would not make it permitted: the fictional prohibition covers private comments themselves, and a comment might contain a recognisable combination of circumstances.
No unauthorised disclosure has been established in the initial facts. Do not accuse the colleague of a breach that has not occurred. If different evidence later shows that private material was entered, follow the organisation’s incident process through the responsible manager; do not promise that deleting a conversation guarantees deletion everywhere.
The federal guide separately discusses information protection and checking how a tool handles inputs. Its federal legal requirements are not being applied to this volunteer organisation. The decision here follows the explicit fictional policy and the need to answer a limited service question without unnecessary disclosure. Government of Canada: information protection.
Find out whose options would actually change
The relevant question is whether people using Tuesday can obtain tools through a workable alternative. Ask the manager to approve a brief consultation through existing communication channels and at both collection windows. Ask about usable times, not medical histories, employment details or reasons people must justify privately.
An open question could be: “If Tuesday collections stopped, which available collection options could you use, if any?” Give a way to respond without discussing personal circumstances in a queue. Do not assume people who say nothing would transfer successfully.
Consultation has limits. Current visitors cannot represent everyone who already finds both windows unusable. A few replies also cannot establish a precise population percentage. Report who was invited, what responses were received and what remains unknown, using the organisation’s authorised information-handling process.
These are original case-specific suggestions. They are not a validated survey instrument or a promise that consultation alone makes a decision fair.
Critique the reassuring shortcut
Consider this response:
I would add a disclaimer saying AI helped write the report, ask the manager to approve it, and reassure visitors that Saturday will be more efficient.
Disclosure identifies the tool’s involvement; it does not repair the reasoning. Manager approval is meaningful only if the manager sees the unsupported assumptions. “More efficient” needs a defined outcome and resource comparison, neither of which the report provides.
The opposite shortcut—discarding everything because AI was involved—also misses useful work. The verified activity counts can remain. The chart can be corrected. A human-written report making the same inference would need the same scrutiny.
A worked recommendation with a decision owner
One defensible response would be:
I would recommend keeping the already funded timetable for next month while we check the proposed change. I would show the manager that both windows recorded three visits per open hour and remove the unsupported claim that everyone can transfer. I would stop the private-comment upload under our policy and propose a limited consultation through approved channels. My revised note would identify the AI-assisted draft, the corrections, remaining uncertainty and the manager’s responsibility for the decision. I would propose reviewing the evidence before the following month’s rota is fixed.
Keeping the timetable has a cost: the existing staffing commitment continues. The supplied facts make that feasible, but do not make it costless. If an urgent staffing shortage emerged, a temporary reduction or different collection arrangement might be necessary. It would need reasons based on that constraint, appropriate notice and a review plan; the flawed AI prediction would not become stronger evidence.
Transfer exercise: accurate wording, missing decision evidence
Original fictional variation. The tool produces an entirely accurate summary of twelve voluntary survey responses. Ten respondents prefer Saturday, two prefer Tuesday, and several Tuesday visitors did not respond. A colleague says, “A human checked every sentence, so the majority proves we should close Tuesday.”
What remains unresolved?
Reasoned feedback. Sentence accuracy does not resolve who is represented, what “prefer” means or whether a less-preferred option is usable. Ten out of twelve describes these respondents, not necessarily all visitors. Ask whether the two Tuesday preferences indicate inconvenience or loss of access, without demanding private explanations.
Keeping both windows while gathering information is defensible if resources permit. Considering a limited change could also be defensible if a real constraint emerges and usable alternatives are confirmed. Neither conclusion follows merely from a checked summary or a majority among respondents. The organisation still owns the tradeoff and the follow-up.
Sources and further reading
- Acuity Insights: How to Prepare for Casper — official preparation starting point; this exercise is independent.
- Government of Canada: Guide on the use of generative artificial intelligence — federal guidance on accuracy, information protection and accountable use; cited as a scoped comparison, not volunteer-sector law.