Question: should an NGO test an AI assistant to draft a short operational or grant-report narrative from a permitted evidence pack? This brief informs that decision, not tool purchasing or final report approval. It is one selected historical evidence set, not a systematic review, a current-model benchmark or proof that this workflow is safe. Full-text methods, results and limitations were appraised on October 27, 2026; supplements, raw data and replications were not audited. No NGO field trial was performed here.
S1, bounded writing: Noy and Zhang's March 10, 2023 working paper, not peer reviewed, randomized ChatGPT access for the second of two occupation-specific writing tasks among 444 college-educated professionals. Three blinded graders assessed outputs. It reports 27 versus 17 minutes and higher grades; see printed pages 2-4 and section 2.4. This is the early working paper, not the final Science version. Its exact model checkpoint is unestablished. Randomization supports the task comparison, while control use and differential attrition complicate it. Self-contained writing tasks did not establish source-grounded NGO-report accuracy, privacy or release safety. Its editing-related grade result does not justify removing verification. Interpretation: drafting usefulness is worth testing, without promising transferable time or quality gains.
S2, evidence integration: Dell'Acqua and colleagues' published March 11, 2026 paper reports a June-2023 GPT-4 experiment, not a 2026 model test. Among 758 consultants, 373 did business analysis from interviews and financial data. On that deliberately challenging task, correctness was 84.5% without AI versus 60%/70.6% in the two AI conditions; see sections 3-4 and Table 7, printed page 11. Other tasks benefited. Randomized comparisons are useful, but one firm's selected tasks, company involvement and incomplete preregistration limit transfer. Interpretation: supplied documents and training do not establish reliable integration. This is neither an NGO error rate nor evidence to stop every drafting use. Do not substitute estimates from the older working paper.
S3, source position: Liu and colleagues' November 20, 2023 v3 full text, later published in TACL in 2024, varies relevant-passage position and distractors across 2,655 questions, and tests synthetic key-value retrieval. Models included GPT-3.5-Turbo 0613, Claude-1.3, MPT and LongChat. See sections 2-5, especially 2.3 and 4.2. Historical 2023 model versions often used information less successfully in the middle; larger context windows did not reliably solve this. Controlled changes help isolate context use, but short-answer matching and synthetic retrieval do not measure NGO narrative completeness, privacy or release outcomes. A query-repetition improvement on synthetic retrieval did not transfer consistently to substantive QA. Interpretation: test omissions under changed pack order; do not prescribe a universal prompt fix or claim current-model error rates.
These are editorial proposals, not validated sufficient controls or an automatic pass score. A study about writing speed does not authorize a report, establish award compliance or clear confidential information for a service.
This is a prepared decision exercise, not an executed AI test. Reuse the grant-change reporting example and its linked canonical fact sheet. Do not create another factual packet or treat its invented labels as research sources.
For a real proposal, attach the AI due-diligence record to the screening note above. Record exact tool/version/date, permitted pack/version, comparator, claim matches, missing required answers, amount/date checks, reviewer effort, failures, owners and the bounded use/hold decision. The organization supplies its own approval criteria; this brief sets no universal acceptable error rate.
The OpenAid site editor owns this one canonical brief; material updates require a separate source/claim reviewer. Record the source version, change reason, review date and resulting use/hold change here rather than replacing the evidence silently or creating parallel briefs. This is a manually reviewed pilot brief. Future reviewer availability and sustained editorial capacity are unconfirmed, so it promises no ongoing update service or review cadence.
Recheck before relying on it after a material task/model/provider change, a source correction or contradictory result, or a failed local check, and by April 27, 2027 if still in use. If a trigger arrives before review is available, hold the affected recommendation and retain the dated evidence and unknowns. Do not infer that an unchanged page means current product approval. No automatic feed, newsletter or subscription collection is provided.
October 27, 2026, first brief: three selected full-text appraisals added. The accessible early S1 version is labeled explicitly; S2 uses the published 2026 version of a historical experiment; S3 addresses supplied-context use. A citation-fabrication study with inaccessible full text was excluded from the appraisal set. Decision remains bounded fictional practice and a hold on unapproved real-data use or unsupported report release.
