Check an AI Draft Against Your Facts

Practice separating a fluent draft from the evidence behind it, then copy a reusable checking prompt.

WhatsAppEmail

The Practical Check

A readable AI answer can contain invented details. Check its factual statements against your supplied material and verify any outside references before using them. Keep personal and sensitive information out of practice inputs. This exercise uses fictional operational facts, and the answer is judged against those facts alone.

Your Starting Facts

Fictional supply-coordination brief: the team has ordered 40 storage boxes. Delivery is expected on November 8 but has not been confirmed. Two staff members will check the delivery. No recipient numbers, cost savings or program results have been supplied.

Worked Example

AI draft: We have delivered 40 storage boxes, helping 600 families and cutting supply costs by 20%. Two staff members checked the delivery on November 8.

Try the Judgment Exercise

Which revision is supported by the brief?

Use the Checking Prompt

Copy this prompt into your own AI tool with non-sensitive inputs, then check the answer yourself. OpenAid does not send this text to an AI service.

Continue With the Resource

Use AI Without Inventing Facts

Optional Facilitator Note

Optional group transfer: after checking the box-delivery draft, open the existing three-record source-appraisal example. Ask participants to decide whether a fluent draft may call L2 independent confirmation or treat H1 as a correction to L1. Use the supplied fictional records only; do not collect real pupil information.

Work Product: A three-row source-fit note labeling original evidence, repetition and changed coverage, followed by one bounded sentence for the planner and one unresolved question with a proposed owner. Keep the record beside the canonical example rather than inventing another fact sheet.

Debrief and Reasons

  1. Does L2 add independent confirmation?

    No. Its later publication repeats L1 without new collection. Trace it to L1 and remove the unsupported area-wide out-of-school claim; publication count is not measurement count.

  2. Can H1 replace L1 or show improvement from 60% to 80%?

    No. One record is center/day physical attendance among enrolled pupils; the other is reported any-school attendance during a different week among children in interviewed households. Denominator, period, method and overlap differ or are unknown. Keep both with their limits; do not average or pool them.

  3. What makes the learner note adequate?

    It limits the L1 statement to 60 of 100 enrolled pupils present at the center on October 12, records H1 separately, and assigns a coverage or definition question to the relevant record owner. Better wording alone cannot supply a missing frame, verify a cause or authorize transport support.

Source and Scope

  • NIST Generative AI Profile

    Sections 2.2 and 2.4; action MS-2.5-003. Supports the factual-output and privacy checks. The fictional exercise and prompt are OpenAid editorial adaptations.

Fictional examples for practice. Completing an exercise does not certify competence.