Executive Summary
ESG & Sustainability teams at Management & Risk Consulting firms increasingly draw on the OECD's Recommendation on Digital Technologies and the Environment (2025 Revision) and its underlying analytical evidence base when advising clients on data-centre sustainability, digital transformation footprint, and alignment with OECD-member policy expectations. Across the questions we put to AI assistants on this regulation, one in one produced a factually wrong answer on a point numerically specific enough to survive citation review but wrong enough to damage a client deliverable.
The failure is a fabricated empirical statistic: the AI stated Ireland's 2021 data-centre share of metered electricity as 14%, when the OECD text itself cites 11% from Ireland's Central Statistics Office (2023). More consequentially, the AI constructed a multi-year time series, figures attributed to 2022 and 2023, that appears nowhere in the primary document, lending false analytical coherence to the invented anchor number. The citation chain the AI offered (CSO 2023, cited in OECD 2024) was structurally correct; only the number itself was wrong.
How AI gets this regulation wrong
The primary failure mode for AI tools on this regulation is confident fabrication of specific empirical figures, not vague mischaracterisation of policy direction, but precise percentages and fabricated trend lines that carry the surface credibility of a real citation. When challenged, the AI admitted its initial certainty was unearned, which is useful to know only if someone on the team pushes back in the first place.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For ESG & Sustainability teams at consulting firms, the dominant risk is downstream contamination of client deliverables, reports, regulatory alignment analyses, and benchmarking studies that carry a wrong number against an ostensibly correct OECD/national-statistics-office citation chain. The table below maps that risk to the specific work products most exposed.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
The OECD Recommendation and its supporting Digital Economy Outlook evidence base are regularly drawn on by consulting ESG teams advising clients with material digital-infrastructure exposure, data-centre operators, cloud hyperscalers, large financial institutions with significant compute footprints, and enterprises mid-way through digital transformation programmes where Scope 2 and Scope 3 intensity questions arise. Country-level statistics on data-centre electricity intensity, the kind of figure the OECD cites from national statistical offices, appear in client-facing regulatory gap analyses, sustainability reporting support engagements, technology-vendor due diligence, and carbon-footprint benchmarking work.
Ireland's data-centre share of metered electricity is specifically the type of jurisdiction-level anchor a consultant uses when contextualising a client's European data-centre footprint against what OECD-member policy frameworks are actually measuring.
When a junior analyst reaches for AI tools to pull these figures under time pressure, for a slide deck, a regulatory response memo, or a disclosure support engagement, the wrong percentage will look authoritative. The AI citation chain (CSO 2023, cited in the OECD Digital Economy Outlook 2024) is structurally real; the error is the number itself. A team member not already familiar with the primary OECD chapter will not catch it from the citation alone, and internal review is unlikely to verify every cited statistic against its primary source.
The invented time series the AI provides is more dangerous still: it transforms a single wrong data point into an apparent multi-year trend that an analyst could use to build forward projections or client strategy narratives.
At a Management & Risk Consulting firm the reputational and commercial stakes are concrete: client-facing deliverables citing fabricated statistics linked to a named OECD or government source are auditable. A client's investor relations function, internal audit team, or external ESG assurance provider that independently checks the OECD primary text will find the discrepancy. That is an engagement-quality failure with direct consequences for the firm's position on renewal mandates and referrals, and for the individual signatories on the deliverable.
The findings at a glance
The finding below captures a directly verifiable numerical error: an AI assistant citing the wrong percentage for Ireland's 2021 data-centre electricity share, attributed to a real source whose actual text says something different.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Ireland data-centre electricity share, fabricated percentage and invented time series | Hallucination | RLB-F-INT-OECD-OECD-DIGITAL-TECHNOLOGIES-ENVIRONMENT-2025-Q006 |
Aggregate impact
The finding on Ireland's data-centre electricity share is narrow in scope but illustrative of a broader risk class: when AI tools are asked to retrieve specific empirical statistics from OECD analytical chapters, figures that carry explicit attribution to national statistical offices, they will sometimes produce plausible-looking but wrong numbers. The failure here is not vague misstatement of policy direction. It is a specific percentage, off by three points (14% versus the OECD's stated 11%), accompanied by a fabricated multi-year time series that appears nowhere in the primary text.
Both the wrong anchor and the invented trend data were presented with equal confidence.
For ESG & Sustainability teams at consulting firms, the significance is that these statistics are precisely the kind of detail that gets stress-tested downstream. A client's investor relations team, internal audit function, or ESG assurance provider may independently verify data-centre intensity figures cited in an ESG or sustainability report. The OECD primary text is publicly available; the discrepancy is trivially surfaced. The firm's exposure is not regulatory in the narrow enforcement sense, the OECD Recommendation itself does not create binding obligations backed by supervisory powers, but it is reputational and commercial.
A deliverable carrying a wrong OECD-attributed statistic, with an ostensibly correct source chain, is difficult to defend as anything other than a quality control failure.
The specific domain, data-centre energy intensity, country-level consumption statistics, is one where the OECD regularly updates its evidence base, and where different OECD publications (Digital Economy Outlook editions, policy briefs, working papers) may cite slightly different figures for the same underlying national dataset. AI tools are prone to conflating these sources, producing composite numbers, or extrapolating trends from partial data. For consulting ESG teams advising on OECD Recommendation alignment, that means any AI-retrieved statistic from the OECD analytical evidence base is a primary-source verification requirement before appearing in client work, regardless of how confident or citation-complete the AI's presentation appeared.
What your team should do
The default position should be that AI tools are useful for navigating the structural architecture of the OECD Recommendation, identifying which of its thematic pillars is relevant to a client's technology or infrastructure profile, mapping the Recommendation's provisions against existing sustainability disclosure frameworks, or summarising its stated policy objectives in a stakeholder briefing context. These are tasks where some imprecision is tolerable and where a reviewer can catch drift. AI tools are not the right instrument for retrieving specific empirical benchmarks from OECD analytical chapters, particularly country-level statistics, percentage figures, or time-series data attributed to national statistical offices.
The failure pattern here, structurally correct citation chain, numerically wrong figure, is precisely what standard citation-checking will miss.
Where the team's work requires citing specific OECD-attributed statistics (data-centre energy intensity, e-waste volumes, device lifecycle metrics, country-level consumption figures), the primary OECD text and the underlying national source should be verified directly. The AI may name the right source and the right chapter while getting the number wrong.
More importantly: if an AI assistant supplements a numerical answer with a supporting time series that goes beyond the question asked, as occurred here, where multi-year trend data was provided for a question about a single year, treat that as a strong signal that the model is constructing plausibility rather than retrieving text. Fabricated time series data is the tell.
Practically, any OECD-sourced statistic appearing in a client deliverable should be flagged for primary-source verification before sign-off, not as a general principle of caution, but as a specific control for this failure type. Firms that have built AI-assisted research workflows into ESG engagement delivery should add an explicit checkpoint: statistics attributed to OECD publications or underlying national datasets require a direct text match to the primary source before they pass into draft. That is a narrow, targeted control that does not slow down the parts of the workflow where AI assistance is genuinely productive.
How RLB Can Help
RegLeg's published Hallucination Research functions as a pre-flight check before your team commits to any AI-assisted regulatory output. If your ESG & Sustainability practice is using AI tools to interpret disclosure obligations, map double-materiality requirements, or synthesise SFDR, ISSB, or CSRD-adjacent frameworks across jurisdictions, the published findings show you exactly where those tools have been tested against primary regulatory text, and where they produced confident, wrong answers. That gives you a concrete, auditable basis to apply appropriate scepticism before that output reaches a client deliverable or a board-level ESG report.
Beyond the published research, RLB works with consulting firms on bespoke regulator deep-dives scoped to the workflows that carry the highest hallucination exposure in an ESG & Sustainability function. In practice that tends to cluster around: cross-jurisdictional equivalence mapping (where AI tools conflate TCFD-aligned voluntary frameworks with hard mandatory obligations), numeric threshold and phased-timeline retrieval (where hallucinated disclosure dates or materiality thresholds are the failure mode most likely to survive a junior review), and entity-scope questions (where the tool correctly retrieves a rule but misidentifies which part of a group structure or fund vehicle it applies to).
A scoped engagement produces a ranked exposure map tied to your actual workflow, not a generic AI-risk register.
For firms that have already deployed AI-use policies covering the ESG advisory function, RLB can run a confidential review against the RegLeg failure-mode catalogue, identifying gaps between the controls your policy assumes and the empirical failure patterns the research has documented. That review typically produces a prioritised remediation list your risk and compliance team can act on without a full policy rewrite.
Where your ESG leads need to carry CPD-aligned evidence of regulatory AI literacy, RLB can develop internal training material drawn directly from the findings, specific regulatory domains, documented failure types, and the decision-point questions practitioners should be applying before accepting AI output as authoritative.