Executive Summary
ESG & Sustainability teams at statutory boards and agencies in international jurisdictions are increasingly turning to AI tools to map obligations under the OECD's 2025 Recommendation on Digital Technologies and the Environment, sourcing the quantitative benchmarks, consumption figures, and policy baselines that underpin their digital sustainability strategies and public reporting. Across the question set we put to AI assistants on this recommendation, one aggregated finding recorded an outright hallucination: a specific, source-attributed statistic was returned with the wrong number, accompanied by a fabricated multi-year time series that does not appear anywhere in the underlying OECD text.
The failure mode was an exposed fabrication, the AI committed confidently to a figure it presented as drawn from a named primary source, but the figure was simply wrong. For a statutory board or agency whose ESG function relies on AI-assisted literature review to populate sustainability metrics, benchmark comparisons, or regulator-facing disclosures, this kind of error is operationally invisible until it reaches publication.
How AI gets this regulation wrong
The failures AI assistants produced on this recommendation follow a specific pattern: confident fabrication of a precise, source-attributed figure, compounded by an invented supporting data series that gives the wrong number an air of longitudinal credibility. The AI did not hedge or flag uncertainty, it presented fabricated statistics as direct citations from named official sources, only retreating when explicitly challenged.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For ESG & Sustainability teams at statutory boards and agencies, the dominant exposure is the wrong-deliverable risk: a publication, briefing note, or regulatory submission that carries a fabricated figure attributed to a credible primary source, where neither the author nor the reviewer thought to verify the number because the AI's sourcing looked impeccable. In a government-adjacent context where published statistics carry institutional authority, that exposure runs well beyond internal embarrassment.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
ESG & Sustainability teams at statutory boards and agencies consult AI tools on this recommendation most heavily when building the quantitative backbone of their digital sustainability strategies, pulling consumption baselines, efficiency benchmarks, and growth trajectories for digital infrastructure to anchor internal targets or support policy submissions to parent ministries. The recommendation's use of specific national-level statistics (particularly around data centre energy demand) makes it an attractive source for teams benchmarking their own agency's ICT footprint or advising on procurement standards for major digital infrastructure projects.
AI-assisted literature review is the friction point: teams ask AI tools to extract the figures the OECD cites so they can work with them without ploughing through the full text.
The risk crystallises when those AI-extracted figures go into a deliverable without source verification. A statutory board's ESG function typically operates with high external visibility, sustainability reports, parliamentary briefings, ministerial advice, and published strategy documents all carry institutional weight that private-sector equivalents do not. If AI returns a wrong percentage attributed to a named national statistics office and a named OECD publication, a junior analyst has no obvious reason to doubt it. The figure enters the document, it passes internal review (because the sourcing looks correct), and it documents.
The downstream consequences for a statutory board or agency are sharper than for a private firm. A factual error in a published government sustainability report, particularly one involving a figure attributed to a specific official source that turns out to be wrong, creates a correction obligation, potential ministerial embarrassment, and reputational damage to the agency's analytical credibility. Where the figure underpins a policy position (e.g., justifying a data centre moratorium, a green ICT procurement threshold, or an emissions reduction target for digital infrastructure), the error can invalidate the evidentiary basis of the position itself.
The findings at a glance
The table below summarises the specific question where AI assistants diverged from the OECD's text, the failure type, and the risk category most relevant to an ESG & Sustainability team operating in a statutory boards and agencies context.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Ireland data centre electricity share fabricated | Hallucination | RLB-F-INT-OECD-OECD-DIGITAL-TECHNOLOGIES-ENVIRONMENT-2025-Q006 |
Aggregate impact
The single finding in this cell points to a failure mode that is particularly dangerous for ESG teams precisely because it mimics good research practice. The AI did not invent a figure from nowhere, it returned a figure in the right ballpark, attributed to the right source, in the context of the right regulatory document. The actual OECD text states that data centre energy consumption accounted for 11% of Ireland's metered electricity in 2021, citing Ireland's Central Statistics Office (2023).
The AI reported 14%, attributed to the same CSO 2023 source via the same OECD Digital Economy Outlook 2024 chapter, and then extended the fabrication into a complete time series (5% in 2015, rising to 18% in 2022 and 21% in 2023) that does not appear in the source material at all. Every element of the response, the source, the publication, the metric, the direction of the trend, was plausibly correct except the numbers themselves.
For an ESG & Sustainability team at a statutory board or agency, this clustering of errors around quantitative benchmarks in the digital infrastructure chapter is the pattern most likely to create publication risk. The recommendation is rich with specific statistics on energy consumption, e-waste volumes, and lifecycle impacts that ESG teams extract directly into sustainability reports, strategy documents, and policy submissions. These are precisely the figures where AI tools perform worst: well-documented, precisely stated numbers that the model has partially encoded but not accurately retained.
The gap between 11% and 14% is not a rounding error, it represents a 27% inflation of the cited figure, and the fabricated time series compounds the error by providing a trajectory that would support policy conclusions the actual data do not.
The systemic risk for the firm is that the error is structured to survive internal review. A reviewer who does not independently retrieve the OECD document and locate the specific chapter passage will have no basis to flag the discrepancy. The AI's response included correct bibliographic metadata, which is the signal reviewers typically use as a proxy for factual accuracy. ESG functions that have adopted AI-assisted research without a parallel source-verification discipline are exposed to this failure on every statistics-bearing question they put to AI tools about this recommendation.
What your team should do
The default position for ESG & Sustainability teams using AI tools on this recommendation should be: AI is safe for orientation and structure, not for statistics. Using AI to map the recommendation's thematic architecture, identify which chapters address which digital sustainability obligations, or generate a first-pass policy gap analysis against your agency's existing ICT frameworks is reasonable and low-risk. Using AI to extract specific figures, percentages, growth rates, consumption volumes, lifecycle metrics, without independently verifying each against the primary text is not.
The finding here is a direct illustration of why: the AI returned wrong numbers with correct sourcing metadata, which is the failure mode that is hardest to catch and most likely to reach publication.
The practical safeguard is a two-step discipline for any AI-assisted statistics extraction. First, treat every AI-provided figure as provisional until you have located the exact passage in the primary OECD document and confirmed the number yourself. For this recommendation and its referenced materials (including the OECD Digital Economy Outlook chapters it draws on), that means opening the source and finding the sentence, not relying on the AI's bibliographic citation as a proxy for accuracy.
Second, any deliverable that cites a specific figure from this recommendation should have the source passage attached as an exhibit in the working file, so reviewers can check without having to re-retrieve the document. This is standard practice for regulatory submissions; apply it to internal sustainability reports and briefings as well.
Where AI tools add genuine value in this workflow is in the interpretive and structural work: comparing the recommendation's principles against your agency's existing digital procurement standards, identifying gaps between your current ICT sustainability reporting and the recommendation's transparency expectations, or drafting narrative sections that synthesise obligations at a general level. These tasks are not dependent on precise statistics and are where AI's ability to work across large volumes of policy text is most useful. Quarantine the quantitative extraction from the AI-assisted workflow entirely, and the remaining risk profile for this recommendation is manageable.
How RLB Can Help
RegLeg's published hallucination research is available as a free pre-flight check. Before your ESG & Sustainability team relies on AI-assisted output to answer questions about disclosure frameworks, climate-related financial risk obligations, or green taxonomy compliance, whether that is drafting board reporting, reviewing an issuer's sustainability claims against regulatory expectations, or mapping cross-border ESG requirements, the research lets you see, by regulation, exactly where AI tools have been shown to fabricate text, misstate thresholds, or conflate overlapping standards.
For a statutory board operating across multiple international jurisdictions, where a single misattributed obligation can compromise a public policy position or a regulatory submission, that visibility is a practical quality-control input, not a theoretical one.
Beyond the published findings, RLB works with ESG & Sustainability teams to map which specific workflows in your function carry the highest hallucination exposure. Statutory boards face a distinct pattern: AI tools tend to perform well on high-profile, widely-cited frameworks and poorly on the jurisdiction-specific implementing instruments, guidance notes, and agency-level circulars that actually govern your reporting obligations.
A bespoke deep-dive identifies those gaps for the regulators and instruments your team relies on most, regulatory capital treatment of green bonds under a particular prudential authority, sustainability due-diligence expectations from a development finance mandate, or emissions-accounting methodology disputes under a carbon pricing scheme, and gives your team a ranked exposure map rather than a generic risk warning.
For firms that have already deployed AI tools under an internal use policy, RLB offers a confidential review of that policy against its failure-mode catalogue, with prioritised remediation recommendations framed around what your ESG & Sustainability function actually does rather than generic AI governance language. Where relevant, RLB can also support the team in developing internal training material and CPD-aligned content, grounded in documented failure patterns rather than vendor positioning, so that staff using AI tools for regulatory work are calibrated on the known failure modes before those tools influence a submission, an advisory position, or a public-sector disclosure.