Executive Summary
ESG and sustainability teams at telecommunications firms in international jurisdictions turn to the OECD's 2025 Recommendation on Digital Technologies and the Environment as a primary reference point for benchmarking their own environmental footprint against the broader digital sector, including the energy trajectory of data infrastructure they operate or depend on. When AI tools were put to questions about the quantitative figures embedded in this Recommendation, the single aggregated finding produced a hallucination: a confidently stated statistic that diverged from the verbatim OECD source and was accompanied by a fabricated multi-year time series that appears nowhere in the document.
The failure mode is particularly sharp because the AI initially committed to its invented figure with full source attribution, and only retreated when directly challenged. For an ESG function whose deliverables are anchored to cited, auditable data, that pattern, confident fabrication followed by retraction, is the worst possible failure profile.
How AI gets this regulation wrong
The failure pattern on this Recommendation is one of fabricated precision: AI tools returned specific numeric figures with full source attribution that do not match the verbatim text of the document, and supplemented those figures with invented trend data that has no basis in the cited source. The breakdown table below maps where that confident-but-wrong dynamic surfaces across the question set.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For ESG and sustainability teams at telecommunications firms, the dominant risk from AI failures on this Recommendation is a wrong deliverable, a report, briefing, or regulatory response that carries a fabricated statistic attributed to a legitimate OECD source. The impact table below traces where that wrong-deliverable risk lands across the team's workflow.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
Telecommunications firms are among the most directly implicated organisations in the OECD's digital-environment framework, they operate the network infrastructure, data centres, and energy-intensive transmission equipment that the Recommendation explicitly targets. ESG and sustainability teams at these firms regularly consult the Recommendation when building out their environmental strategy, calibrating Scope 2 and Scope 3 reporting boundaries, or positioning the firm's footprint against sector-level benchmarks.
AI tools get pulled into that workflow at the point of research: a team member asks for the headline statistics the OECD cites to substantiate the framework's urgency, the kind of figure that ends up in a board sustainability deck, an annual ESG report, or a response to a regulatory data call.
The specific failure mode here, an AI returning a plausible but wrong percentage for data-centre energy consumption, complete with a fabricated multi-year time series, is acutely dangerous in that workflow because the error is invisible at first glance. The AI attributed its figure to the correct sources (CSO 2023, OECD Digital Economy Outlook 2024), presented a coherent narrative, and offered no uncertainty signal. A junior analyst building a benchmarking slide would have no obvious reason to go to the primary document.
That figure then travels forward into the ESG report, into peer-comparison analysis against other European telcos, and potentially into a regulator's data collection exercise, all with a fabricated number that overstates the sector's energy burden by roughly 27% relative to the actual cited figure.
The stakes are compounded by the Recommendation's status as an OECD Council-level instrument: national regulators in multiple jurisdictions use its data as a baseline for domestic policy design and enforcement expectations. An ESG team that inadvertently misquotes an OECD benchmark in a regulatory submission is not just correcting an internal document, it is potentially explaining to a national authority why its published environmental data diverges from the source the authority itself relies on.
The findings at a glance
The table below summarises the one aggregated finding from AI testing on this Recommendation that is directly relevant to ESG and sustainability work at telecommunications firms.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Fabricated Ireland data-centre energy share and time series | Hallucination | RLB-F-INT-OECD-OECD-DIGITAL-TECHNOLOGIES-ENVIRONMENT-2025-Q006 |
Aggregate impact
The single finding on this Recommendation is narrowly scoped, one quantitative figure in one country-level example, but the failure pattern it reveals is structurally significant for ESG teams at telecommunications firms. The OECD Recommendation relies heavily on cited statistics to establish the environmental urgency that motivates its policy asks. Those statistics are precisely the type of content ESG teams lift into their own documents: benchmark figures, sector comparisons, growth trajectories. When an AI tool fabricates a percentage and then invents a plausible-looking time series around it, the error does not look like an error, it looks like additional supporting data.
The finding clusters on quantitative benchmarking within the energy-consumption narrative of the Recommendation, which is the section most likely to be consulted by a telecommunications ESG team building a regulatory-environment section for a sustainability report or a board briefing. The OECD's verbatim text puts Ireland's 2021 data-centre share of metered electricity at 11%, with a 144% growth rate between 2015 and 2020. The AI tested produced 14% and fabricated a four-point time series (5%, 14%, 18%, 21%) across 2015–2023, figures that do not appear in the source but are internally coherent enough to pass a light editorial review.
The systemic risk is one of citation laundering: an authoritative source (OECD/CSO) is real, the document is real, the topic is real, only the number is wrong. For a telecommunications firm operating under mandatory ESG disclosure requirements in multiple jurisdictions, a fabricated OECD statistic in a published report is not a minor data-quality issue. It is a potentially material misstatement in an audited document, traceable to a verifiable primary source that says something different.
What your team should do
The default position for any AI-assisted lookup of cited statistics in this Recommendation should be source verification before the figure leaves the research stage. That means going to the primary document, the OECD Digital Economy Outlook 2024 chapter that contains the data-centre energy figures, or directly to the CSO 2023 publication the OECD itself cites, before the number appears in any work product. AI tools are useful for identifying which section of the Recommendation contains the relevant data and for flagging the cited sources; they are not reliable for returning the verbatim figure.
The specific failure pattern here, confident attribution to a real source with a wrong number, will not trigger standard editorial review because the source citation checks out on the surface.
In practice, the workflow safeguard is a two-step rule for any quantitative OECD figure entering an ESG deliverable: AI identifies the passage and its citation chain; a team member retrieves and checks the primary source before the number is used. This is especially important for figures that have acquired a second life in policy discourse, energy-consumption percentages for data centres are frequently cited, frequently rounded, and frequently updated, which creates exactly the conditions in which AI tools confuse versions or interpolate trend data that was never published.
The Ireland 11% figure is a live example of this: it appears in a specific snapshot year (2021) and should not be extrapolated into a time series without direct primary-source support.
AI tools are reliably useful in this workflow for navigating the structure of the Recommendation itself, identifying which of its six sections and associated annexes address the energy-intensity of networks and data infrastructure, mapping it to analogous domestic regulatory instruments across the firm's operating jurisdictions, and drafting the framing prose around verified figures. The risk is concentrated specifically in the retrieval of cited quantitative data from embedded references within OECD publications; everything else in the ESG team's use of this Recommendation remains a reasonable AI-assisted task provided outputs are reviewed by someone who knows the document.
How RLB Can Help
RegLeg's published Hallucination Research is a pre-flight check your team can run before placing weight on AI output for any regulatory question, disclosure thresholds under CSRD sector-specific standards, jurisdictional scope determinations for TCFD versus TNFD alignment, or transition plan requirements that vary materially between the EU, UK, and APAC frameworks your group entities face. The research is structured around the failure modes that matter in practice: wrong entity, wrong number, wrong obligation, inverted position.
Before your team routes a gap-analysis query or a materiality assessment through an AI tool, checking whether that regulation appears in RegLeg's catalogue tells you in advance where the model has a demonstrated track record of getting things wrong, and what category of error it tends to produce.
Where the published research doesn't cover your specific regulatory stack, we run bespoke regulator deep-dives scoped to your workflows. For a Telecommunications ESG & Sustainability function the highest-exposure workstreams are typically: AI-assisted disclosure drafting against cross-jurisdictional reporting standards (CSRD, SDR, MAS ESG guidelines), supply chain due diligence obligation interpretation (CSDDD applicability to non-EU group entities, network infrastructure supply tiers), and spectrum or licence-condition sustainability commitments that sit at the intersection of sector regulation and ESG policy.
We map which of those workflows carry the greatest hallucination risk, by failure mode and by the specific regulatory instrument, so your team knows where to hold the line on AI-assisted output and where it is safe to use as a first draft.
For firms that have already deployed AI tools internally, we can review your existing AI-use policy against our failure-mode catalogue and return a prioritised remediation list, grounded in findings against the specific regulations your team cites most often, not generic AI governance principles. We can also convert that analysis into training material and CPD-aligned content your team can use directly: scenario-based exercises built from real failure examples drawn from the regulatory instruments relevant to your reporting obligations, formatted for internal delivery without exposing methodology or proprietary research detail.