What Has the World Bank Announced?

In a publication dated 8 October 2026, the World Bank Group explained how it is testing AI-assisted legal research for its Building Green dataset. The dataset examines building energy codes and their enforcement, helping researchers understand both what regulations require and how those requirements operate in practice.

The next edition covers 119 cities across 110 economies. For the legal research component, an AI workflow generates answers to 60 questions for each city. Each answer must be linked to the relevant legal source, including an exact excerpt and a URL that researchers can inspect.

OpenAI GPT-5.4 Mini was used to draft answers, while Anthropic Claude independently reviewed the legal basis in supervised sessions with an analyst. Human researchers also checked the answers against the underlying sources.

The initiative demonstrates a practical use of AI in public-sector research: helping researchers process large collections of legal documents without treating generated answers as automatically reliable.

What Is the Building Green Dataset?

Building Green is a World Bank Group dataset focused on building energy codes and their enforcement. Building energy codes establish requirements intended to improve energy efficiency in buildings, particularly through standards that apply to construction and related building practices.

Such regulations can help governments establish minimum energy performance requirements. However, the existence of a legal requirement does not necessarily mean that it is consistently implemented or enforced.

The dataset therefore combines legal research with information from local experts. This distinction helps researchers examine not only what regulations say, but also whether implementation arrangements support compliance.

How Is AI Being Used to Research Building Regulations?

The World Bank workflow is designed to identify relevant laws, retrieve legal texts and connect each answer to its supporting evidence. The process prioritizes official legal sources rather than relying on general summaries.

What Do the Initial Accuracy Results Show?

The World Bank reported that approximately four in five answers were correct across the first 21 economies reviewed. Accuracy was at least 75% in 18 of those 21 economies, reaching as high as 95% in the strongest-performing economy.

The system performed particularly well on questions with clear yes-or-no or fixed-choice answers. Nearly nine in ten answers in these categories were correct, according to the publication.

These results are preliminary and do not establish that every answer across all 119 cities is correct. The World Bank identified remaining challenges in determining the latest amendments to legal codes and identifying all project types and exemptions covered by specific rules.

The findings support a cautious approach: AI can accelerate research, but accuracy must be assessed against authoritative legal documents and the requirements of each jurisdiction.

Why Legal Accuracy Matters in AI Research

Building regulations can involve several layers of legislation. A general law may establish a framework, while a detailed ministerial regulation or technical standard specifies the actual energy efficiency requirements.

An AI system may identify a relevant rule but cite the wrong legal instrument. It may also rely on an outdated regulation that remains accessible online even after it has been amended or replaced.

The World Bank describes three important safeguards: checking the legal source, independently reviewing the legal basis and requiring analysts to verify the result. The workflow also allows the system to return an insufficient-evidence response when the available information does not justify a confident answer.

These safeguards are particularly important when research results may inform public policy, regulatory comparisons or decisions about building energy standards.

AI Research Costs and Time Savings

The World Bank reported substantial reductions in research time during its testing and production work from March to August 2026. Previously, a researcher could need approximately two to three days per jurisdiction to identify relevant legal instruments, read them and prepare 60 answers with legal references.

With the AI-assisted workflow, reviewing the drafted answers for one city takes approximately four to six hours. The publication also reported about US$3,200 in OpenAI usage costs across the 119-city project, or roughly US$27 per city.

The project produced a benchmark of 7,140 human-graded rows, covering 60 questions across 119 cities. This benchmark can help researchers evaluate future automated verification systems using reviewed examples rather than assuming that an AI system is reliable.

These figures describe the reported project and its specific workflow. They should not be interpreted as guaranteed cost savings or accuracy levels for every legal research project.

Why Building Energy Regulations Matter for Climate Action

Buildings and construction account for around 37% of global carbon dioxide emissions, according to the World Bank publication. Building energy codes are among the policy tools governments can use to establish energy efficiency requirements for new construction.

Understanding these regulations across jurisdictions can help researchers compare legal requirements, identify implementation gaps and examine where enforcement may be inconsistent.

The AI project does not itself introduce new building standards or establish a global climate regulation. Instead, it aims to improve the collection and verification of information about existing national and local requirements.

What Comes Next?

The next edition of the Building Green dataset is expected in spring 2027. It is intended to present legal requirements alongside information from local experts about how those requirements are applied in practice.

Comparing the two sources can reveal important differences. A legal requirement may exist on paper while implementation is inconsistent because of limited inspection capacity, shortages of qualified professionals or difficulties obtaining compliant materials.

The planned dataset could help researchers and policymakers identify questions that require further investigation. The final findings will depend on the completed research, source verification and expert input; the planned release should not be treated as already published.

What Other Public-Sector Researchers Can Learn

The World Bank project offers a model for using AI in research based on public records. Its approach separates questions that can be answered from published documents from questions that require professional experience or observations of real-world implementation.

• Use authoritative primary sources for factual and legal claims.
• Attach exact source excerpts and URLs to generated answers.
• Allow an insufficient-evidence response rather than forcing an unsupported conclusion.
• Use independent verification and human review.
• Track recurring errors and improve the workflow with explicit validation rules.
• Evaluate performance using human-checked examples and documented accuracy results.

These principles can be useful in other research projects involving public regulations, government records and policy datasets. Their effectiveness still depends on the quality of the sources, the design of the questions and the rigor of the review process.

Frequently Asked Questions

Conclusion

The World Bank Groups AI-assisted Building Green research project shows how AI can support large-scale regulatory research when paired with exact legal evidence, independent review and human verification. The work covers 119 cities in 110 economies and is intended to contribute to a broader understanding of building energy codes and their implementation.

Its central lesson is that AI can make document-heavy research more efficient, but credible results require authoritative sources, transparent evidence and careful checks. The next Building Green dataset, expected in spring 2027, will provide further information about how building energy regulations operate in practice.