AI-assisted development reporting: separate activity evidence from outcome claims

Decide which question the update can answer

Start with the decision the reader needs to make. A quarterly donor update may need to say which activities occurred, where records conflict and what evidence is still required. It need not assert that an intervention improved livelihoods. The OECD's development evaluation criteria distinguish relevance, coherence, effectiveness, efficiency, impact and sustainability. They are lenses for questions, not a ready-made method or permission to claim impact from a dashboard. Select only the criteria useful to the actual update and state what evidence would be needed for stronger conclusions.

For AI-assisted work, ask a second question: whose account is missing? The UNDP's account of the Hamburg Declaration on responsible AI describes commitments to inclusion, human rights and local innovation. A central-office summary should therefore not erase local context merely because the English-language spreadsheet is easier to process. That principle informs the exercise below; the declaration is not a validation of this particular method or a claim that any programme achieved its goals.

Sources: [1] [2]

Build a claim-to-evidence sheet

Imagine a fictional programme that ran digital-service workshops in three districts. The synthetic pack contains an attendance register, an English progress slide, a partner's local-language note and an earlier target table. The progress slide says “all districts completed the planned workshops.” The register lists sessions in two districts and a blank entry for the third. The partner note says the third session was postponed because the venue was unavailable. An AI summary that repeats the slide as fact has missed the decisive contradiction.

Make one row per proposed statement with columns for claim, intended audience, source document and date, translation status, corroboration, responsible local reviewer, confidence and permitted wording. The first row can say that two sessions are documented by the register. The third-district row is not “completed”; it is “completion unconfirmed, with a partner note reporting postponement.” A separate row may record that participants attended, but it cannot become “participants gained lasting digital capability” without an appropriate follow-up measure. Keep activity, immediate output and outcome in different rows.

Use AI only to suggest a first extraction of names, dates and discrepancies from approved, non-sensitive material. The worker compares every extracted datum with the originals. If the local-language note was machine translated, preserve the source wording and translation status, then ask a qualified local colleague to confirm the meaning before publishing an explanation. Do not infer consent to share a quotation from its presence in a field note. Remove identifying details from a public update unless the programme's own permissions and safeguarding process allow them.

Verify before writing the narrative

Run three checks on the sheet. First, reconcile counts: can the attendance register, session calendar and target table describe the same period and unit? Second, test interpretation: does a local reviewer agree that “postponed” describes the third event, or is the translation ambiguous? Third, test the strongest sentence: if the draft says the workshops improved access to services, what observation or comparison supports that claim? If the answer is only attendance, narrow the sentence to an observed activity and name the outcome question for later measurement.

The reviewer should be able to trace every sentence in the update to a row. A useful final note might say: “The available register documents two workshops this quarter. The third district's status requires confirmation from the partner before this report is finalized. Attendance does not establish whether participants later used the service.” That is less dramatic than an AI-generated success story, but it is actionable: the programme owner knows what can be reported, what to verify and who should resolve the gap.

Keep programme and assessment boundaries visible

This is a synthetic evidence-handling exercise, not advice about a specific country, donor requirement or safeguarding law. A real programme must apply its own agreements, data permissions, language review and local accountability. Two Prune could use a bounded simulation to observe whether a participant catches the slide/register conflict, preserves the local note's uncertainty and writes a calibrated update. Its report would describe task-specific assessment evidence for human interpretation; it would not certify an intervention's impact or make a funding or employment decision.