Executive Summary
Across the OECD's 2025 Recommendation on Digital Technologies and the Environment, we tested AI tools on the quantitative and evidentiary claims that Professional Engineers in international jurisdictions are most likely to reach for, infrastructure energy benchmarks, jurisdictional data sourced from national statistics bodies, and figures that carry citation chains back to primary regulators. On one aggregated question, AI tools returned a fabricated answer: a specific statistical figure attributed to a named primary source that did not match the verbatim text of that source.
The failure pattern here is not vagueness or omission, it is confident numerical fabrication. The AI tool produced a precise percentage for Ireland's 2021 data-centre share of metered electricity, attributed it correctly to Ireland's Central Statistics Office via the OECD Digital Economy Outlook 2024, and then, when pressed, retracted the figure. It also generated an accompanying multi-year time-series, internally plausible, stylistically authoritative, entirely absent from the underlying document.
For Professional Engineers whose deliverables depend on traceable, citation-verified statistics, this failure mode carries direct professional exposure: a fabricated figure in an engineering report, an environmental impact assessment, or a policy brief looks identical to a real one until it is checked against the primary source.
How AI gets this regulation wrong
The dominant failure mode AI tools exhibit on this regulation is confident numerical fabrication: presenting a specific, source-attributed figure that differs from the verbatim text in the underlying document, accompanied by invented corroborating data that makes the error harder to detect. What makes this pattern particularly acute here is that the AI tools involved did not hedge, they named the source, stated the figure with precision, and only walked it back when directly challenged on the derivation.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your practice
For Professional Engineers, the risk this failure creates is a wrong deliverable: a report, opinion, or submission that enters the record carrying a misattributed statistic and a fabricated supporting series. The table below maps where in practice that exposure lands, and how visible the error is likely to be before the work product leaves your hands.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects Professional Engineers
Professional Engineers working on digital infrastructure, data centre siting, grid interconnection studies, lifecycle assessment of ICT equipment, or environmental compliance for hyperscale operators, regularly need to ground their work in current jurisdictional consumption benchmarks. The OECD Recommendation and its accompanying Digital Economy Outlook chapters are go-to citation sources precisely because they aggregate national-statistics-body data across member countries and present it with a credible multilateral imprimatur. Engineers reach for AI tools to retrieve those figures quickly: what share of national grid does the data-centre sector represent in a given country, and what has the trajectory been?
That is exactly the query type where the failure documented here surfaces. An engineer drafting a technical annex for an environmental impact assessment, scoping a data-centre project against Ireland's CSO-reported baselines, or preparing a policy submission that references OECD-cited national figures would receive a precise-sounding percentage, 14% in the tested instance, attributed with apparent accuracy to CSO 2023 via the OECD publication. The actual verbatim figure is 11%. The AI also supplied a multi-year trend series (5% → 14% → 18% → 21%) that has no counterpart in the document.
Both the point figure and the series would pass a casual review; neither would survive a check against the source text.
The professional exposure is sharper than it might first appear. Engineering reports typically enter formal processes, planning applications, grid operator submissions, regulatory consultations, where cited statistics are scrutinised by counterparties. A figure that deviates 3 percentage points from the primary source, or a fabricated trend series, is the kind of error that surfaces during technical review or expert cross-examination and reflects on the engineer's verification practice, not just their choice of research tool.
The findings at a glance
One aggregated finding was identified against this regulation, a question on a specific citation-attributed statistic where AI tools produced a confidently stated but incorrect figure, with no surface indication of uncertainty until challenged.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Ireland 2021 data-centre electricity share, fabricated figure and invented time-series | Hallucination | RLB-F-INT-OECD-OECD-DIGITAL-TECHNOLOGIES-ENVIRONMENT-2025-Q006 |
Aggregate impact
With a single aggregated finding in this cell, there is no cluster pattern to map, but that single finding is instructive precisely because of what it reveals about where AI tools fail on this regulation. The OECD Digital Technologies and Environment framework is unusual in that its authority rests heavily on its citation infrastructure: the recommendations derive credibility from the national-statistics-body data they aggregate. An AI tool that gets those underlying figures wrong, and presents the error as a verbatim citation, undermines the very thing that makes the document useful as an engineering reference.
The fabrication here was not a loose paraphrase. The AI tool committed to a specific percentage (14% vs. the verbatim 11%), attributed it to the correct source chain, and then generated a corroborating multi-year time-series that had no basis in the underlying text.
This is the compound failure that matters for Professional Engineers: the incorrect point figure would have been catchable by anyone who pulled the primary source, but the fabricated series is more insidious, it looks like contextual enrichment, the kind of surrounding data that makes a figure feel anchored, and it would not be detected by a reviewer who did not independently run the primary source.
For engineers working across multiple OECD jurisdictions, where similar consumption-share figures exist for Germany, the Netherlands, Singapore, and other data-centre-dense markets, the implication is systemic. If AI tools fabricate with this level of specificity on the Ireland figure, the same failure mode will manifest on comparable national-statistics figures for other jurisdictions. Any workflow that uses AI to retrieve jurisdiction-specific benchmarks from OECD-cited sources, without a primary-source verification step, carries this exposure across every country covered.
What your team should do
The default position on AI-retrieved citation statistics in OECD publications should be: treat every specific figure as unverified until you have read the sentence it appears in from the primary source. That is not a counsel of paranoia, it is the same verification standard that applies to any secondary reference in a technical report. The fabrication pattern documented here is precisely the kind that a fast-moving junior would trust and a careful reviewer would catch too late; building the primary-source check into the workflow upstream of drafting is cheaper than finding it downstream.
Where AI tools are safe on this regulation is in orientation and scoping: identifying which chapters of the Digital Economy Outlook address a particular jurisdiction, flagging which OECD working-party outputs are relevant to a given technology domain, or summarising the structure of the recommendation's implementation guidance. These are tasks where an approximate answer is useful and a precise figure is not at stake.
Similarly, AI can usefully draft the framing sections of a technical annex, the regulatory context, the policy objectives, the cross-jurisdictional comparison structure, as long as the specific numbers that populate those sections are entered from primary sources by the engineer, not by the AI.
For teams that regularly work on data-centre or digital-infrastructure mandates across OECD jurisdictions, the practical safeguard is a citation verification checklist specific to OECD-sourced statistics: every figure that traces back to a national statistics body (CSO, INSEE, CBS, etc.) cited via an OECD publication must have the OECD source text alongside it, not just the figure. If the AI cannot produce a verbatim excerpt containing that figure, the figure is unverified, regardless of how precisely it is stated or how well the citation chain is described.
How RLB Can Help
RegLeg's published Hallucination Research is available as a free pre-flight reference for Professional Engineers before relying on AI output for regulatory questions. The research catalogues specific failure modes, misquoted thresholds, outdated standards references, fabricated cross-citations, observed across the regulatory instruments that govern engineering practice in multiple jurisdictions. Reviewing findings relevant to a regulation before using AI tools on a matter saves time and reduces the risk of acting on output that sounds authoritative but does not reflect current requirements.
For firms with multiple Professional Engineers working across the same regulatory portfolio, RegLeg offers bespoke regulation deep-dives. These engagements examine the specific instruments your team relies on, map the failure modes most likely to surface in your practice context, and produce a structured reference your engineers can apply consistently. The output is practical and jurisdiction-specific, designed to sit alongside existing compliance workflows rather than replace professional judgement.
RegLeg also develops training material and CPD-aligned content tailored to the failure modes Professional Engineers should actively watch for when AI tools are part of a regulatory research workflow. Where a firm already has an AI-use policy in place, RegLeg can provide a confidential review of that policy against its failure-mode catalogue, identifying gaps, ambiguous guidance, or areas where the policy does not adequately address the known ways AI tools mishandle technical regulatory content.