Evidano home
Published by the World Bank Independent Evaluation GroupResponsible AIWorld Bank Group

When Can Evaluators Trust AI?

Evidano6 min read
Estelle Raimondo, Head of Evaluation Methods, Data and Analytics at the World Bank’s Independent Evaluation Group

An evaluation team can now put thousands of papers, reports, or interview transcripts within reach of a large language model. The temptation is obvious: search faster, summarize more, and classify evidence at a scale that would be difficult to manage manually. The harder question comes afterward. When an AI system produces a plausible answer, what evidence is enough to trust it?

That is the problem behind “Balancing innovation and rigor: Guidance on how to thoughtfully integrate AI in evaluation,” a 2025 article by Harsh Anuj, Estelle Raimondo, and Hannah Den Boer. Drawing on experiments at the World Bank’s Independent Evaluation Group (IEG) and IFAD’s Independent Office of Evaluation, the authors argue that AI should be treated as a component in a designed and validated workflow, not as a shortcut around methodological rigor.

We spoke with Estelle Raimondo, Head of Evaluation Methods, Data and Analytics at the World Bank’s Independent Evaluation Group, to learn more about how evaluators can test where AI adds value without weakening standards of evidence.

About the article

Research question

How can evaluators use large language models to process and analyze large volumes of text while preserving analytical rigor? The article asks how teams should select use cases, design experiments, measure performance, and decide what counts as acceptable performance for a particular task.

Key finding

AI usefulness should be tested task by task rather than assumed. The authors recommend representative sampling, separate validation and testing, context-specific performance thresholds, iterative prompt refinement, and mandatory human review.

Why this research matters

Evaluators often work with more text than a team can examine exhaustively by hand. In one application described by the authors, the workflow was used to identify, bulk download, and screen the full text of more than 10,000 research papers. That creates a real opportunity for AI, but also a risk: speed can make a process look comprehensive without proving that the output is complete, relevant, or suitable for the decision at hand.

The article turns “responsible AI” into operational questions. What task is the model performing? What errors matter most? What human-labelled material will serve as a benchmark? Which metrics fit the task, and what threshold is acceptable given the stakes? Raimondo reflected on the broader significance of this shift: “There are many normative frameworks out there for responsible AI in evaluation, the next frontier is to be able to put these norms into practice. We have found that starting from the workflow design makes things tangible and operational; helping us to put discourse into action.”

This matters beyond development evaluation. Researchers using AI for literature reviews, evidence synthesis, document classification, or large-scale text analysis face the same distinction: a fluent output is not the same as a validated output, and a threshold that works for early screening may be inadequate for a judgement that affects policy or program decisions.

Key findings from the article

Three contributions stand out because they show how the authors move from enthusiasm about AI to testable standards.

  • Classification performance has to be interpreted through the task, not a single headline score. In the literature-identification experiment, GPT-4o achieved 0.90 accuracy, 0.75 recall, 0.60 precision, an F1 score of 0.67, and balanced accuracy of 0.67 on a 30-paper test set. The team accepted this trade-off because missing relevant evidence was more costly than manually checking extra false positives.
  • Generative outputs could be faithful yet still incomplete. For 30 generated paper summaries, relevance averaged 4.87 out of 5 and coherence 4.97, with no hallucinations observed under the study’s faithfulness measure, although some incorrect aggregation occurred. Information extraction, by contrast, scored only 3.25 for relevance despite perfect faithfulness scores on the measures used. An answer can avoid inventing facts and still omit information that matters.
  • Rigor came from the validation loop around the model. The authors split data into training, validation, testing, and prediction sets, used representative examples, and required manual review when LLMs or machine learning were used. Their proposed strategy uses roughly five representative examples for training, around 20 for validation, and around 30 for testing, with subject-matter experts providing labels and disagreements examined during refinement.

Taken together, the findings suggest that the hardest part of AI-assisted evaluation is establishing the conditions under which an output deserves evidentiary weight. Raimondo reflected on the results: “IEG’s AI strategy puts ‘quality first’ as its mantra for the scale up of AI. We take it seriously and invest heavily into experimenting and testing properly. This means that human review, validation, manual labeling and bias checks is embedded in the journey, not just at the end but throughout.”

Methods and research approach

The underlying guidance note reports two months of on-the-job experiments in late 2024 by a multidisciplinary IEG and IFAD team. The main use case was a structured literature review for an ongoing IEG evaluation of World Bank Group support for epidemic preparedness. Rather than automating an entire review, the team broke the workflow into tasks such as identification, extraction, annotation, summarization, and synthesis.

Four components were tested. Classification used GPT-4o on a 30-paper test set. Summarization used GPT-4o-mini on 30 papers. Synthesis generated a 500-word account from six 200-word summaries. Information extraction used 12 papers and returned 57 scored responses about actors, mechanisms, and goals. Prompts were refined on validation material before evaluation on held-out test data.

The limits matter. Only classification was embedded in a live evaluation use case; the generative tasks were experimental and lacked a human-generated benchmark. The authors did not test the full review or synthesis workflow, individual experiment samples were modest, and their ability to test larger open models was limited. The work is therefore best read as a practical framework from applied experiments, not a universal benchmark of LLM performance.

Broader implications

For evaluation teams, the article suggests starting with the task, not the model: where is current practice inefficient or too shallow, what added value could AI provide, and how would success be measured? That framing makes it easier to reject weak use cases and to demand stronger evidence when the stakes are higher.

Human oversight is also part of the method, not a final check after automation. Subject-matter experts are needed to create labels, review edge cases, define acceptable error, interpret disagreements, and decide when AI output can support an evaluation claim. The authors also recommend shared annotation guidelines and intercoder reliability measures so the human benchmark is examined rather than treated as unquestionable ground truth.

The remaining challenge is standardization. The authors call for more experimentation across organizations and contexts. Raimondo reflected on what should come next: “While it is hard to resist the hype and speed of AI development, there is value in going slow and steady, experiment thoroughly, test before you scale, repeat use cases across contexts, and co-create guidance so they are actually owned and enforced. It seems to me that we will be in this phase of the journey for quite a bit longer.” The wider implication is measured but consequential: AI can expand the evidence evaluators can examine, but credibility still depends on transparent design, explicit validation, and human judgement.

About Estelle Raimondo

Estelle Raimondo, Head of Evaluation Methods, Data and Analytics at the World Bank’s Independent Evaluation Group

Estelle Raimondo is Head of Evaluation Methods, Data and Analytics at the World Bank’s Independent Evaluation Group, where she advises and trains teams across a wide range of methodologies, spearheads methodological innovation, and leads the adoption of data science and AI in evaluation practice. Her research includes evaluation methods, causal analysis, the practical integration of data science into evaluation, and the institutionalization of evaluation.

About the authors

  • Harsh Anuj (Independent Evaluation Group, World Bank Group, United States)
  • Estelle Raimondo (Independent Evaluation Group, World Bank Group, United States)
  • Hannah Den Boer (Independent Office of Evaluation, International Fund for Agricultural Development, Italy)

Read the article

Read the full research article, “Balancing innovation and rigor: Guidance on how to thoughtfully integrate AI in evaluation,” published by the World Bank Independent Evaluation Group in 2025.

Full citation

Anuj, H., Raimondo, E., & Den Boer, H. (2025). Balancing innovation and rigor: Guidance on how to thoughtfully integrate AI in evaluation. Independent Evaluation Group, World Bank Group. ieg.worldbankgroup.org/blog/balancing-innovation-and-rigor-guidance-how-thoughtfully-integrate-ai-evaluation

View the publication

We are grateful to Estelle Raimondo for sharing her perspective on this work and for her contribution to the responsible, transparent use of AI in evaluation.

Research, accuracy and responsible AI

How Evidano's output has been evaluated, how its models were built, how your data is handled, and how to disclose or cite AI use in your own work.

  • Resource

    Human vs AI: validated accuracy benchmarks

    Four head-to-head comparisons against expert human analysis — 371 interview transcripts with Arizona State and Penn State, 298 evaluation reports, 224 open-ended survey responses with UT Austin and Wichita State, and a 160-response manual-coding review.

  • Resource

    How Evidano’s models were trained

    Model provenance for research-ethics review: a provenance-tiered curated corpus, and customer uploads, prompts and outputs excluded from training entirely.

  • Resource

    Data security for qualitative research

    Encryption in transit and at rest, customer-controlled deletion, regional storage, and SOC 2 Type 2, GDPR, HIPAA-ready and CCPA/CPRA compliance.

  • Resource

    How to cite and disclose Evidano

    Citation formats in APA, MLA, Chicago, Harvard and Vancouver, methods-section wording per workflow, and an AI-policy review of 41 journals across 8 disciplines.