Executive Summary
For ESG & Sustainability teams at Electricity & Power firms operating across international jurisdictions, the OECD Recommendation on Digital Technologies and the Environment (2025 Revision) carries direct relevance: it establishes policy expectations around digital infrastructure energy use, lifecycle impacts, and the green transition obligations governments are expected to embed in national frameworks. When those teams turn to AI tools to interrogate the Recommendation's specific empirical anchors, the country-level figures and citations that ground disclosure narratives and regulatory mapping, the AI produces confidently wrong numbers.
Across the question set tested against this regulation, AI tools fabricated a statistic central to the Recommendation's justification for digital infrastructure governance, attributing a specific percentage to an OECD-cited national source while the actual figure in the text is materially different. The failure mode is particularly hazardous because the AI not only misquoted the primary figure but invented an entire supporting time-series, creating a coherent-looking but wholly fabricated data trail that a junior analyst would have no obvious reason to question.
How AI gets this regulation wrong
The failure pattern on this regulation is one of confident fabrication: AI tools stated a specific empirical figure with apparent precision, attributed it correctly to the primary source, and then, when tested further, acknowledged they were not certain of the underlying number. What makes this particularly dangerous for teams working with OECD soft-law instruments is that the AI's invented statistics look indistinguishable from genuine citation work, complete with source attribution and trend narrative.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For ESG & Sustainability teams in the power sector, the risk from this finding lands squarely in the wrong-deliverable category: a disclosure, policy brief, or regulatory submission that carries a fabricated OECD-cited statistic into an external audience. The exposure is not abstract, power firms increasingly use OECD digital environment benchmarks to contextualise their own data-centre and grid-digitalisation footprint disclosures, and a materially wrong reference figure can compromise both the credibility of the filing and the team's standing with regulators who hold the primary text.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
ESG & Sustainability teams at power firms encounter this Recommendation most directly when they are building the regulatory context sections of climate and digital transition disclosures, particularly where board-level reporting needs to situate the firm's data-centre offtake, grid digitalisation capex, or smart-meter rollout against the policy environment OECD member governments are expected to operationalise.
The Recommendation's empirical anchors, national-level data-centre energy shares, lifecycle impact figures, cross-sector digital-green nexus statistics, are exactly the kind of third-party references that a team will pull to justify why a given disclosure topic is material, or to respond to a regulator's question about the firm's exposure to emerging digital infrastructure obligations.
The same team will also reach for AI assistance when scoping how the Recommendation's provisions interact with jurisdiction-specific implementation: which national regulators have incorporated its digital energy efficiency expectations into guidance, what the baseline consumption figures imply for grid operators versus end-users, and where the Recommendation's language overlaps with mandatory reporting frameworks the firm already files under. All of these use cases involve the AI retrieving or synthesising specific figures and source attributions, precisely the task where the failure documented here occurred.
If the AI's answer travels unchecked into a board sustainability report, a regulator-facing policy submission, or a due-diligence briefing on a data-centre PPA counterparty, the firm carries a materially incorrect OECD-cited statistic into a permanent record. In international jurisdictions where regulators are beginning to scrutinise digital infrastructure energy disclosures against the very OECD benchmarks the Recommendation references, a wrong number creates both a factual error to retract and a process question about the reliability of the team's research workflow.
The findings at a glance
The table below summarises the finding tested against this regulation, the type of AI failure observed, and its risk impact category for ESG & Sustainability teams at Electricity & Power firms.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Ireland data-centre electricity share, fabricated figure and time-series | Hallucination | RLB-F-INT-OECD-OECD-DIGITAL-TECHNOLOGIES-ENVIRONMENT-2025-Q006 |
Aggregate impact
The single finding from this regulation reveals a failure pattern that is characteristic of OECD soft-law instruments more broadly: the Recommendation's empirical content is drawn from secondary citations to national statistical offices, and AI tools are poorly calibrated on the precise figures those citations carry. The AI tested here did not simply estimate or hedge, it committed to a specific percentage (14%) that diverges materially from the primary text (11%), then constructed a multi-year time-series around the fabricated anchor to give it narrative coherence.
The result is a response that reads as well-sourced and internally consistent while being factually wrong on the number that matters.
For ESG & Sustainability teams in the power sector, this failure clusters on exactly the kind of content the function uses most: benchmark figures that contextualise the firm's own energy footprint relative to the digital infrastructure economy. Data-centre electricity consumption shares, grid digitalisation energy intensities, and lifecycle impact multipliers are the raw material of materiality assessments, scenario analyses, and regulatory context sections. An AI-generated figure that is wrong by three percentage points on a jurisdiction the regulator has explicitly cited is not a rounding error, it is a mis-statement of the regulatory record.
The systemic risk to the firm is compounded by the AI's behaviour when challenged: in testing, the AI acknowledged uncertainty only after the initial confident delivery, meaning a team that does not probe the answer will never encounter the hedge. Junior analysts following a research workflow that treats AI as a first-pass lookup are the most exposed, because the failure is invisible at the surface of the response and the fabricated time-series provides false corroboration that would satisfy a normal plausibility check.
What your team should do
The default position for this regulation is straightforward: treat AI tools as incapable of reliably retrieving specific statistics from OECD-cited national sources, and build a mandatory primary-source check into any workflow that outputs a number attributed to the Recommendation or its underlying data chain. The OECD text is publicly available; the relevant chapter and the CSO citation it carries can be verified in minutes.
Any disclosure, policy brief, or regulatory submission that references data-centre energy shares, digital infrastructure lifecycle figures, or other empirical anchors from this Recommendation must have those figures verified against the primary document before sign-off, not as a belt-and-suspenders step, but as a non-negotiable control.
Practically, the team should treat AI assistance on this regulation as safe for structural and thematic tasks: mapping the Recommendation's provision structure, identifying which national implementation frameworks reference it, drafting narrative framing around obligations the team already knows, or generating initial outlines for policy gap analyses. These uses do not depend on AI accurately retrieving specific figures, and the failure documented here does not affect them.
The risk concentrates on any task where the AI is being asked to supply a number, percentage, or citation, particularly where the underlying source is a national statistical office report mediated through an OECD secondary citation.
Where the team's workflow involves benchmarking the firm's digital infrastructure footprint against OECD member-country data, for example in PPA due diligence on data-centre counterparties, or in climate transition plan sections addressing the digital-energy nexus, the control is to retrieve the relevant OECD chapter directly and extract figures from the primary text. Building a short reference sheet of the Recommendation's key empirical anchors (with chapter, paragraph, and cited source noted) is a low-cost safeguard that removes the retrieval task from AI entirely and prevents the fabricated-time-series failure mode from entering any team output.
How RLB Can Help
RegLeg's published Hallucination Research gives your team a concrete pre-flight check before you rely on AI-assisted output for regulatory work, climate disclosure obligations, taxonomy alignment assessments, scope 3 reporting thresholds, just-transition requirements. The findings are regulation-specific and publicly accessible, so your team can query them directly when an AI tool returns a confident-sounding answer on a piece of regulation you're about to act on. That's not a substitute for your own read of the primary document, but it is a systematic record of exactly where the tools fail on the texts you're likely working with.
Where the published research is the starting point, the bespoke work goes further. For ESG and sustainability functions in electricity and power, the highest hallucination exposure tends to cluster around the intersections regulators get wrong themselves, delegated acts still in consultation, taxonomy criteria that differ between jurisdictions with otherwise identical labels, and disclosure thresholds that have been amended without the amendments appearing cleanly in the consolidated text.
RegLeg can map your specific AI-supported workflows, CSRD/ESRS gap analysis, EU Taxonomy do-no-significant-harm screening, grid infrastructure permitting disclosures, carbon market compliance, against its failure-mode catalogue to show you where the tools are most likely to return authoritative-looking errors and where the risk of acting on those errors is material.
If your firm already has an AI-use policy in place, RegLeg can review it against the failure patterns documented in the research and return a prioritised remediation list, specific gaps in your guardrails rather than a generic framework audit. For teams building internal capability, RegLeg can produce CPD-aligned training material grounded in real hallucination examples from the regulatory texts your team works with, which tends to land differently with practitioners than vendor-supplied AI literacy content that wasn't written with electricity sector ESG obligations in mind.