Executive Summary
ESG and sustainability teams at digital platforms and marketplaces firms operating across international jurisdictions rely on the OECD's 2025 Recommendation on Digital Technologies and the Environment to anchor their environmental measurement frameworks, data centre energy disclosure commitments, and supply-chain climate reporting obligations. When those teams turn to AI tools to interrogate the empirical backbone of this Recommendation, the specific statistics and cited data points that underpin policy positions, the results carry material errors.
Across the question set tested against this regulation, AI tools got the facts wrong in ways that produce authoritative-sounding but incorrect figures: fabricated percentage statistics, invented time-series data, and misattributed citations presented with confidence and precision. A team that ingests these outputs into a regulatory mapping document, a CDP submission, or an investor-facing ESG report will embed a verifiably false data point sourced to a real OECD publication, a combination that is both credible-looking and wrong.
How AI gets this regulation wrong
The dominant failure pattern AI tools exhibit on this regulation is confident fabrication of quantitative data: AI tools state specific figures, percentages, growth rates, multi-year time series, that do not appear in the underlying OECD text, attribute them to real sources, and initially defend them as accurate before walking them back under challenge. The table below maps out where this occurs and which types of factual claims are most at risk.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For ESG and sustainability functions at digital platforms and marketplaces firms, the primary exposure from AI failures on this regulation falls in the wrong-deliverable category: work products, disclosures, benchmarking analyses, policy briefs, that carry a false factual claim presented as an OECD-sourced figure. The table below shows where in the ESG workflow these errors convert from an AI mistake into a firm-level risk.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
ESG teams at digital platforms and marketplaces firms use this Recommendation most intensively when benchmarking their own data centre energy performance against international reference points, the OECD figures on national-level data centre electricity consumption are precisely the kind of contextual data that goes into CDP disclosures, investor ESG questionnaires, and the environmental sections of annual or integrated reports. Teams also reach for this regulation when mapping their firm's voluntary environmental commitments against OECD member-state policy trajectories, or when scoping the regulatory risk horizon for new infrastructure builds in jurisdictions where OECD guidance is incorporated into national frameworks.
The second common use-case is internal: sustainability analysts producing briefings for the board, for investor relations, or for cross-functional product teams need accurate empirical anchors for claims about the digital economy's aggregate environmental footprint. When a junior analyst asks an AI tool to retrieve the specific figure for data centre electricity consumption in a given country as cited by the OECD, they expect a verbatim lift from the primary source, not a plausible-sounding approximation.
The difference between 11% and 14% of Ireland's metered electricity, to take the concrete case, is not rounding noise: it is a 27% overstatement that, if published, is directly contradicted by the OECD document the team is citing as authority.
If that error reaches an investor-facing disclosure or a regulatory submission, the firm has a verifiably false claim in a document where accuracy is a legal and reputational obligation. For digital platforms operating across multiple OECD-member jurisdictions, the exposure compounds: incorrect baseline figures distort the internal carbon-accounting benchmarks that inform target-setting, capital allocation for efficiency investments, and supplier due-diligence thresholds. Remediation requires identifying every downstream document that inherited the figure, issuing corrections, and in some disclosure frameworks, notifying regulators, a process that is expensive, visible, and entirely avoidable.
The findings at a glance
The following table summarises each tested question, the AI outcome, and the failure type, giving your team a direct read on where AI tools broke down against the specific text of this regulation.
| # | 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 error pattern in this regulation's findings clusters tightly on quantitative citation retrieval, specifically, AI tools fabricating the precise statistics the OECD uses to justify its policy positions and framing those fabrications as direct source citations. The Recommendation's empirical case for digital infrastructure governance rests on numbers like data centre electricity consumption shares, energy growth rates, and country-level environmental intensities.
These are exactly the figures ESG teams need to get right, and exactly where AI tools perform worst: a figure cited as 11% in the primary source becomes 14% in the AI response, accompanied by a fabricated multi-year trend that sounds analytically coherent but has no basis in the document.
The systemic risk for digital platforms and marketplaces firms is structural rather than incidental. These firms tend to have disproportionately large digital infrastructure footprints relative to their headcount, which means data centre energy intensity is often their most material environmental metric. When ESG teams benchmark against OECD reference figures, for internal target-setting, for investor engagement, or for national-level regulatory compliance mapping, an inflated baseline figure does not just affect one disclosure: it cascades through every calculation that uses it as an input.
A wrong baseline figure in a 2025 Scope 2 benchmark can distort 2026 reduction targets, misrepresent performance against peers, and create a discrepancy that becomes visible only when a sophisticated investor or regulator compares the firm's cited source against the primary OECD document.
The compounding factor is the confidence with which AI tools present these fabricated figures. In the tested case, the AI did not hedge or flag uncertainty, it cited the specific OECD publication and the specific national statistics office as the authority for a figure it had invented. That framing actively discourages the source-verification step that would catch the error. Teams operating under time pressure, closing a disclosure, preparing for an investor call, responding to a regulator query, are precisely the teams least likely to double-check a figure that arrives pre-cited with an authoritative provenance trail.
What your team should do
The default position for ESG and sustainability teams at digital platforms and marketplaces firms should be: AI tools are not a safe retrieval mechanism for specific quantitative figures cited in OECD publications. The failure mode identified here, confident fabrication of a precise percentage, complete with attribution to a real source, is not detectable from the AI response itself. The only reliable safeguard is primary-source verification for every specific statistic before it enters any work product.
This means maintaining a discipline of going directly to the cited OECD chapter or the referenced national statistics office publication and confirming the figure verbatim, treating any AI-supplied number as a research lead rather than a citable fact.
Where AI tools remain genuinely useful in this regulation's workflow is at the structural and directional level rather than the numerical level: identifying which parts of the Recommendation are most relevant to a given use case (data centre governance, circular economy for hardware, climate disclosure alignment), drafting policy-mapping frameworks that will be populated with verified data, or summarising the regulatory intent behind specific provisions in plain language for internal briefings. These tasks do not require AI to retrieve precise figures, they require reasoning about structure and scope, where the error risk is substantially lower.
For teams building out their AI-tool governance for ESG workflows, this regulation is a useful calibration case. The OECD Recommendation is a public document with specific, verifiable empirical claims embedded in it. Test any AI tool your team uses against a known figure from this text before relying on it for quantitative retrieval.
If the tool cannot return the correct 11% figure for Ireland's 2021 data centre electricity share without prompting, it should not be in the workflow for any quantitative ESG benchmarking task involving OECD source material, and that test should be repeated each time the tool or its underlying model is updated.
How RLB Can Help
RegLeg's published Hallucination Research is available as a free pre-flight check for ESG & Sustainability teams before committing to AI-assisted regulatory analysis. When your function is working through disclosure obligations, supply-chain due diligence frameworks, or taxonomy alignment questions, the research flags precisely where AI tools have mischaracterised scope, misstated thresholds, or conflated overlapping regimes, the failure modes that surface as errors in a sustainability report or a regulator submission, not in a QA review.
Treating the published findings as a standing reference costs nothing and calibrates how much independent verification your team needs to layer on top of AI output on any given question.
Beyond the public research, RLB conducts bespoke regulator deep-dives scoped to the specific AI-supported workflows that carry the highest hallucination exposure for ESG & Sustainability in a digital platforms or marketplace context. Cross-border greenwashing rules, platform-liability carve-outs in sustainable-finance disclosure regimes, and the interaction between data-governance obligations and ESG reporting pipelines are all areas where AI tools have demonstrated systematic failure patterns under test conditions. A deep-dive maps those patterns to your actual workflow touchpoints, so your team knows where to apply scepticism and where AI assistance is reliable enough to accelerate work without added review cost.
RLB also offers a confidential review of your firm's existing AI-use policy against the RegLeg failure-mode catalogue, with a prioritised remediation list that your ESG & Sustainability lead and your compliance function can action jointly. Where the policy has gaps, around AI-generated regulatory summaries, automated materiality assessments, or AI-assisted stakeholder disclosures, the review surfaces them in terms your legal and governance colleagues will recognise. CPD-aligned training material is available alongside the review output, giving your team internal documentation that demonstrates structured engagement with AI risk in the ESG function, useful both for regulator enquiries and for board-level AI governance reporting.