Disclosing the Energy and Water Impact of Artificial Intelligence: Analysing Transparency in Environmental, Social, and Governance (ESG) Reports of Australian Financial Service Companies
- Policy Research Program
- 12 minutes ago
- 17 min read
Authors: Mia Conway and Sophie Mahura
INTRODUCTION
The rapid adoption of artificial intelligence (AI) technologies across the global economy has generated increasing concern regarding the environmental sustainability of AI-related corporate practices and the adequacy of existing frameworks for disclosing them. In Australia, companies have been embedding AI tools into daily operations at an accelerating rate, with early pilot programs justifying broader enterprise-level expansion driven by demonstrated gains in productivity and efficiency (Singla et al. 2025). The environmental costs of this expansion, however, are substantial. The energy and water intensities of the data centre infrastructure that houses AI systems have been extensively documented, positioning AI as a significant and growing threat to national and global climate goals (Vanderbergh et al. 2025). This tension is particularly consequential within the financial services sector, where AI has been rapidly adopted across cybersecurity threat detection, fraud management, risk assessment, and document processing (UTS 2023: 6), and where companies simultaneously carry public commitments to sustainability and net-zero targets. As organisations scale their AI operations, questions emerge regarding whether current reporting frameworks adequately capture and disclose the environmental costs associated with that infrastructure.
Despite growing research on AI sustainability and the evolution of corporate ESG reporting, limited work has examined how Australian financial services companies disclose the environmental impacts of AI within existing sustainability frameworks. This paper addresses that gap by providing a detailed investigation into current disclosure trends among the ten largest financial sector companies by market capitalisation listed on the Australian Stock Exchange (ASX), using their most recent FY2024-25 annual and sustainability reports. The study analyses disclosures through the qualitative data analysis software NVivo to examine how and to what extent companies are transparent about AI's energy and water demands, and how their disclosures align with established climate-related reporting frameworks.
Environmental, social, and governance (ESG) reporting frameworks have gained increasing significance in the corporate governance landscape as climate pressures have transformed stakeholder expectations, requiring greater transparency from companies about their operational impact on the environment (Zervoudi et al. 2025: 5-7). International reporting frameworks have been established to provide wide-scale guidance on what environmental metrics to report and how, with the International Sustainability Standards Board (ISSB) and the Global Reporting Initiative (GRI) representing two of the most widely adopted in practice. Sustainability reporting has evolved from a voluntary practice to a mandatory one across many jurisdictions (Khamisu and Paluri 2024), and in Australia this shift is expressed through the 2024 amendment to the Corporations Act 2001. As of January 2025, large entities that produce financial reports are required to produce a sustainability report that includes climate statements covering material risks and metrics on Scope 1, Scope 2, and Scope 3 GHG emissions, in accordance with the requirements of AASB S2 Climate-Related Disclosures (Australian Securities and Investments Commission 2025). As AI becomes more central to business operations, understanding how companies are already engaging with these standards to disclose AI usage and environmental impact helps to identify gaps in current frameworks and offers guidance for addressing those gaps in future iterations.
The International Sustainability Standards Board's climate disclosure standard, IFRS S2 (2023), and its Australian legislative equivalent, AASB S2 (2023), require entities to organise climate-related disclosure across four interconnected pillars: governance, strategy, risk management, and metrics and targets, which together provide a structured basis for assessing how companies manage and report climate-related financial risks (ISSB 2023: 7). Applied to AI-related energy risk, the framework requires that companies demonstrate board and management oversight of AI's energy demands under governance, show how those demands inform strategic and transition planning under strategy, embed AI energy consumption within formal risk identification processes under risk management, and report quantitative metrics or reduction targets attributable to AI workloads under metrics and targets. For the financial services sector specifically, where significant AI infrastructure is operated through third-party cloud providers, the four-pillar structure creates disclosure obligations that extend beyond direct operational emissions to encompass Scope 3 transition risk exposures. Historically, however, Australian listed companies have exhibited inconsistent and incomplete climate-related disclosures, with physical and transition risks frequently underrepresented in financial reports (Yang et al. 2024: 265), a pattern that provides important context for the findings of this study.
GRI 302 Energy (2016) provides the primary international framework for disclosing organisational energy consumption, with indicators spanning total consumption within the organisation (GRI 302-1), energy consumed outside the organisation including upstream Scope 3 emissions from cloud computing services (GRI 302-2), energy intensity ratios (GRI 302-3), and reductions in energy consumption linked to operational changes (GRI 302-4). This distinction is critical to the findings of this study, since company reports may aggregate total electricity use in ways that satisfy GRI compliance without disaggregating AI-specific or data centre loads, while only the latter demonstrates meaningful transparency about AI's environmental footprint. GRI 303 Water and Effluents (2018) applies analogously to water consumption in data centre cooling, covering withdrawal by source (GRI 303-3), net consumption volumes (GRI 303-5), and interaction with shared water resources (GRI 303-1). Given that high-performance computing facilities used in AI training and inference require significant quantities of water for cooling operations (SC Adopting AI 2024: 146), disaggregated water reporting represents an equally significant yet frequently overlooked dimension of AI-specific environmental disclosure.
Globally, the integration of AI's environmental impacts into sustainability reporting remains fragmented and insufficiently developed, with frameworks such as GRI and ISSB instrumental in harmonising sustainability practices more broadly but not yet updated to incorporate AI-specific disclosure requirements (Mustafa et al. 2025: 9762). The absence of applicable reporting standards makes AI's environmental impacts difficult to quantify, with available data further constrained by commercial sensitivity (SC Adopting AI 2024: 149). The intersection of AI, environmental reporting, and greenwashing risk remains significantly underresearched (Moodaley and Telukdarie 2023: 4), and within Australia, the absence of sector-specific AI emissions reporting frameworks leaves companies operating under significant disclosure discretion, which the literature suggests tends to produce selective and inconsistent reporting practices in contexts where anticipated reputational risks are high (Yang et al. 2024: 266).
METHODOLOGY
The companies analysed in this paper were sampled from the Listcorp Australian Stock Exchange, classifying public companies into industry groups in accordance with the Global Industry Classification Standard (GICS). GICS is a four-tiered, hierarchical industry classification system comprising 11 sectors, 25 industry groups, 74 industries, and 163 sub-industries (MSCI 2023). For this study, the 10 largest financial-sector companies by market capitalisation were selected, encompassing the banks, financial services, and insurance industry groups (See Figure 1).
The primary data source used to analyse company disclosure levels was each company's most recent annual report for the FY2024-25 reporting period, accessed directly from company investor relations pages. Accompanying ESG and sustainability reports were also included where available. As (Yang et al. 2024: 268) note in their comparable analysis of ASX-listed company climate disclosures, annual reports represent the primary venue through which listed companies communicate with investors and are therefore the most appropriate source for assessing the nature and extent of formal disclosure practices. These documents were uploaded to NVivo, a qualitative data analysis software, and coded against a six-node framework developed specifically for this study to extract data and identify thematic patterns across the sample.
The six nodes were structured to move sequentially from recognition to accountability, reflecting the analytical logic of the study's research question. Node 1, AI reference type, captured the manner in which AI was named or implied within each document, distinguishing between AI explicitly named in an environmental context, AI implied via infrastructure references such as data centres, cloud computing and servers, AI referenced in a risk context, and instances where AI appeared in the document without any environmental framing. Node 2, disclosure level, assessed the depth of any AI-related environmental disclosure present, coding outcomes as absent, qualitative only, or quantitative, with quantitative disclosure further distinguished between AI-specific figures and proxy figures drawn from broader digital infrastructure data carrying no AI disaggregation. Node 3, environmental impact category, identified the type of environmental consequence referenced, spanning energy consumption, Scope 2 carbon emissions, Scope 3 carbon emissions, water consumption, and broad environmental impacts where no specific resource category was named. Nodes 4, 5 and 6 benchmarked these disclosures against three established international frameworks. Node 4 assessed alignment with the four pillars of ISSB S2, covering governance, strategy, risk management, and metrics and targets. Node 5 assessed alignment with GRI 302 energy indicators, including disaggregated energy consumption within and outside the organisation, energy intensity ratios, and reductions in energy consumption linked to digital operations. Node 6 assessed alignment with GRI 303 water indicators, including interactions with water as a shared resource, management of water-related impacts, water withdrawal by source, water discharge by destination, and net water consumption. Coding followed a deliberate sequential order, with Nodes 1 through 3 completed before any framework benchmarking was undertaken, to prevent normative judgments about expected compliance from influencing the initial assessment of what was actually disclosed. Interpretive coding decisions were documented in linked NVivo memos attached to the relevant company node to maintain an auditable record of analytical reasoning throughout the process.
While this study seeks to contribute to research supporting greater disclosure of AI's environmental impact, its purpose was to measure disclosure transparency and framework alignment rather than to verify the actual environmental consequences of AI operations across the sampled companies, and the findings do not constitute evidence of verifiable environmental impact. Several methodological limitations bear acknowledgment. The reliance on self-reported annual and sustainability reports as the sole data source means that the findings are bounded by what companies chose to disclose, and cannot account for environmental impacts that exist but were not reported. This constraint is structurally inherent to content analysis methodologies applied to voluntary disclosure contexts, where the object of analysis and the gap under investigation are both constituted by the same reporting practice (Khamisu et al. 2025: 3). The sample size of 10 companies, while appropriate for an in-depth qualitative analysis of the largest entities by market capitalisation, limits the generalisability of findings to smaller financial services companies operating outside the top tier. Sequential coding was adopted as the primary mitigation against interpretive bias, ensuring that framework benchmarking did not retrospectively shape the initial reading of each document, though the inherently interpretive nature of qualitative content analysis means that residual subjectivity in coding decisions cannot be entirely eliminated. NVivo memos were used to document and make transparent the reasoning behind borderline coding decisions, providing a degree of analytical accountability. Inter-coder reliability testing was not undertaken given the constraints of the study's scope and timeline, and this represents a limitation that future research in this area should seek to address.
FINDINGS
The findings reveal a structurally consistent pattern across the sample, where AI is acknowledged as a commercial and operational priority, situated within governance and strategy language that satisfies the declaratory requirements of international frameworks, yet systematically from the quantitative environmental accounting that would render it genuinely transparent.
The most instructive finding concerns the language companies chose when referring to AI in annual reports, since those choices directly determine which environmental consequences become visible in disclosure. The AI implied via infrastructure node returned the highest coding volume of any AI reference type, generating 34 references across 9 sources. This substantially exceeded AI explicitly named in an environmental context, which returned 26 references across 8 sources, and AI in risk context, which returned 17 references across 7 sources. A further 23 references across 4 sources were coded to no AI-environment disclosure, with AI appearing exclusively in commercial, operational or product contexts that carried no environmental framing. The practical significance of infrastructure-implied AI references is that companies routinely disclose data centre operations, cloud computing arrangements and GPU-based processing in ways that gesture toward AI's environmental footprint without explicitly naming it as a driver of energy or water demand. CBA's FY2025 report illustrates this pattern directly. The report details the launch of its AI Factory platform, built on Amazon Web Services infrastructure and capable of processing 3 trillion data points across 4 billion transactions, with GPU-based models performing tasks 90 per cent faster than CPU predecessors, yet the energy implications of this computational expansion do not appear anywhere in the document's sustainability disclosure. A comparable pattern is evident in Macquarie's FY2025 report, which references AI extensively in the context of fraud detection, behavioural biometrics and commercial product development while AI remains entirely absent from environmental framing. Across both cases, and across the sample more broadly, companies emphasise information that creates a favourable impression of AI capability while withholding information that would expose the extent of its environmental impact, a pattern consistent with selective disclosure as a strategic communication practice (Yang et al. 2024: 265).
The consequences of this naming behaviour become apparent when considered alongside the disclosure level findings, which reveal that the dominant outcome across the sample is the complete absence of AI-related environmental disclosure. The absent node returned 37 references across 8 sources, representing the highest volume of any disclosure level outcome. Qualitative-only disclosure accounted for 22 references across 5 sources, capturing instances where companies acknowledged AI's environmental implications in general terms without attaching figures to those acknowledgments. The quantitative proxy node, which captures aggregate data centre or digital infrastructure figures carrying no AI-specific attribution, returned 30 references across 6 sources. The AI-specific quantitative node, representing figures directly attributed to AI workloads, returned zero sources and zero references across the entire sample. This constitutes the most significant empirical result of the study: no financial services company in the sample produced AI-disaggregated quantitative energy or water disclosure in FY2024-25. Where quantitative environmental data exists, it is presented at the level of organisational aggregates that neither identify nor isolate AI as a contributing factor, meaning that investors and regulators cannot currently determine what proportion of any company's energy or water consumption is attributable to AI operations. The environmental impacts of AI are difficult to quantify in part because commercially sensitive data on data centre energy and water consumption remains opaque at an industry level, and the absence of AI-specific quantitative disclosure in this sample suggests that this opacity is being reproduced, rather than resolved, at the individual company reporting level. That no company in the sample has produced AI-disaggregated figures, despite the rapid scaling of AI operations across the sector, indicates that the discretion currently afforded to companies under voluntary frameworks is consistently exercised in favour of aggregation, and that this pattern will persist for as long as such discretion remains.
The environmental impact node data reveals that where environmental disclosure does occur, it gravitates toward categories that least expose the disaggregation gap. Scope 2 carbon emissions returned 21 references across 7 sources and broad environmental impacts returned 11 references across 8 sources, while energy consumption as a named category returned only 8 references across 5 sources, and water consumption returned 7 references across 5 sources. Scope 2 and broad environmental framings are by nature organisation-wide and do not require AI-specific attribution to satisfy formal disclosure conventions. Energy and water consumption, by contrast, are the categories for which GRI 302 and GRI 303 provide the most operationally specific reporting guidance, and for which AI and data centre operations create the most direct and measurable physical demand. Suncorp's sustainability data pack illustrates the resulting gap, disclosing detailed water consumption methodology drawn from building-level metering across 13 monitored sites with quarterly extrapolation applied to unmetered locations, yet linking no portion of that consumption to data centre cooling or AI operations despite AI being actively deployed across the organisation. Medibank's sustainability summary similarly acknowledges that expanded cloud software and licensing have contributed to increases in its Scope 3 GHG emissions, yet frames this as a categorical aggregate rather than an AI-specific disclosure. This allows the environmental consequences of AI adoption to be acknowledged in principle without being quantified in practice, a distinction that carries material consequences for both comparability and accountability.
The ISSB S2 alignment findings are the most revealing of all six nodes, as they expose the relationship between governance language and measurable accountability most directly. A pronounced concentration in the narrative pillars of the framework is accompanied by a sharp deficit at the level of metrics and targets. Energy strategy returned 22 references across 7 sources, governance returned 20 references across 6 sources and risk management returned 12 references across 5 sources, while metrics and targets returned only 7 references across 3 sources, the weakest positive outcome across the ISSB S2 framework. Four sources returned no ISSB S2 alignment across 8 references. This pillar-level distribution describes a sector that is comfortable articulating board-level oversight and strategic intent in relation to AI energy risk while declining to produce the quantitative targets and performance metrics that would allow those declarations to be evaluated. ANZ explicitly acknowledges in its FY2025 annual report that AASB S2 will formally apply from the financial year commencing 1 October 2025, signalling awareness of forthcoming mandatory obligations, yet produces no AI-specific metrics or targets in the current reporting period. QBE discloses a 2030 GHG emissions target described as specific, measurable and time-bound, yet no component of that target is disaggregated or attributable to AI-related activity. The result across the sample as a whole is a form of ISSB S2 engagement that satisfies the declaratory requirements of the governance and strategy pillars without generating the quantitative accountability that the metrics and targets pillar is specifically designed to procure. This pattern will persist for as long as discretion over the depth of disclosure remains with individual companies, since the incentive structures identified across this study consistently favour narrative adequacy over quantitative specificity.
This trend is reinforced by the GRI 302 and GRI 303 findings, where references to framework compliance carrying no AI-specific disaggregation dominate so comprehensively that they constitute the defining characteristic of the sector's approach to energy and water reporting. Under GRI 302, the references but not disaggregated node returned 19 references across 1 source, representing the dominant outcome by a substantial margin, while GRI 302-1, 302-2 and 302-3 returned zero sources and zero references, and only one source generated any references under GRI 302-4. Under GRI 303, the references but not disaggregated node returned 14 references across 10 sources, again the dominant outcome, while GRI 303-1 and 303-2 returned zero references across the sample, and 6 sources were coded to no GRI 303 alignment across 13 references. ANZ reports in accordance with GRI Universal Standards 2021 and publishes a GRI Content Index, yet no AI or data centre load is disaggregated within its energy or water figures. IAG's ESG data summary references GRI frameworks across multiple data tables without AI-attributable figures appearing within them. Macquarie's sustainability report is prepared with reference to GRI, TCFD and SASB simultaneously, yet AI-specific metrics and targets remain absent from its environmental disclosures. Together, these results demonstrate that a substantial gap exists between an organisation citing GRI compliance and an organisation producing disaggregated AI-specific data that would constitute genuine alignment with GRI 302's energy indicators or GRI 303's water indicators. Across the entire sample, no company has bridged that gap. Even where frameworks are cited, Australian financial services companies are not using them to produce the kind of AI-specific environmental accountability that the accelerating scale of AI deployment across the sector demands (Mustafa et al. 2025: 9762).
DISCUSSION
The six-node framework applied in this study was constructed to move systematically from surface recognition to deeper structural alignment, producing findings whose thematic coherence is worth examining in sequence. Coding against Node 1 (AI reference type) revealed that explicit AI naming was frequently decoupled from environmental framing, with companies identifying AI in the context of operational capability or digital transformation while omitting connection to resource consumption. As such, a selective discourse approach was overwhelmingly identified where positive associations with technology are foregrounded and adverse impacts are suppressed (Moodaley and Telukdarie 2023:4). Node 2 (Disclosure level) suggested a strong skew toward qualitative-only or wholly absent disclosure. Where qualitative figures appeared, they typically described broad organisational energy consumption, falling short of the granularity that GRI302-1 specifically requires for meaningful analysis (GRI 302 2016:9). Node 3 (Environmental impact category) showed that where environmental language was present it concentrated on carbon emissions, primarily Scope 2, with water consumption mostly absent. Nodes 4 through 6 then benchmarked these raw disclosures against ISSB 2, GRI 302, and GRI 303 respectively, ISSB S2 alignment, where detectable, was concentrated in governance and risk management pillars, suggests that sampled companies has begun to internalise AI energy risk at a structural level while underreporting the qualitative accountability that ISSB S2 requirements ultimately demand (ISSB S2 2023: 13). GRI 303 alignment was the most consistently absent across the sample, reinforcing the observed pattern that qualitative ESG disclosure methods predominate in sectors where measurement standards remain underdeveloped (Khamisou et al. 2025: 12).
The disclosure pattern documented in this study is best understood as consistent with international findings rather than as a distinctively Australian failing. GRI 302, GRI 303 and ISSB S2 are internationally applied frameworks and the literature suggests that their adoption across multiple jurisdictions has produced similarly incomplete AI-specific environmental disclosures. Across the United States, AI firmed disclosure of only limited energy and environmental information that no jurisdiction has succeeded in targeting disclosure of AI-driven energy use specifically, with even European Union regulatory initiatives remaining unclear in their practical reach (Vanderbergh 2025: 4). This reflects a structural condition common to all contexts in which these frameworks operate where ESG disclosure remains fragmented and subsequently procured inconsistent metrics and prevent meaningful cross-cm[any comparison (Badmus et al. 2025: 3978). The disclosure gap documented in this study is therefore an expression of a globally consistent problem located at the level of the standards themselves (Mustafa et al. 2025: 9775).
Three intersecting factors account for the thin, qualitative-dominance and non-disaggregated disclosure observed across six nodes. First, metering limitations are structural, given that no accepted standards currently exist for attributing energy or water consumption specifically to AI workloads as distinct from general digital infrastructure. The environmental impacts of AI remain difficult to quantify, with energy and water profiles considered opaque due to commercial sensitivity and the absence of standardised reporting requirements (SC Adopting AI 2024: 149). This leaves even motivated disclosers unable to produce figures that meet GRI 302-3 intensity ration or GRI 303-5 net consumption requirements. Second, the voluntary status of AASB S2 during the study period removes the regulatory compulsion that has produced disclosures elsewhere, as the transition from voluntary to mandatory frameworks is consistently the primate driver of ESG disclosure quality (Khanmisu et al. 2025: 12-14). Third, asymmetric incentive structures compound the technical gap where companies hold reputational incentive to publicise AI capability, signalling innovation and digital leadership to investors while disclosure of AI’s energy and water costs carry reputational risk with the same audience (UTS 2023: 18). This asymmetry is self-reinforcing in ways that the findings of this study make visible. Companies that reference AI capability extensively across their annual reports while coding as absence or qualitative-only under Node 2 are responding rationally to a disclosure environment in which stakeholder pressure to report environmental costs remain weak relative to the reputational rewards of signalling digital leadership. Companies tend to calibrate their communications to satisfy the expectations most visible to them and across the sampled companies, those expectations have not yet coalesced around AI environmental performance specifically (Zervoudi et al. 2025: 11). The result, observable in the divergence between nodes, is that disclosure follows incentive rather than impact, with AI capability attracting extensive narrative attention while energy consumption, water use and framework alignment remain the least developed dimensions of reporting across the sample (Mustafa et al. 2025: 9779).
Taken together, these factors raise a question that extends beyond the reach of voluntary GRI compliance alone, namely whether sector-agnostic energy and water standards, applied voluntarily, are sufficient to surface the AI-specific environmental impacts of financial services companies. The findings presented in this study suggest their limitations are material. GRI 302 and GRI 393 are designed to apply to any any organisation regardless of sector (GRI 302 2016: 5, GRI303 2018: 5) and their sector-agnostic architecture confines the granularity of disaggregation they can require as AI-workloads effectively subsumed within the broader totals these standards are made to capture. The absence of uniform materiality thresholds across ESG frameworks limits comparability and undermined investor confidence in what that are directly applicable to the patterns this study has documented (Badmus et al. 2025: 3979) The Australian regulatory trajectory therefore requires, at minimum, a sector-specific guidance on AI infrastructure disclosure within the ASSB S2 implementation schedule. Without such guidance, voluntary GRI alignments will continue to produce the aggregated, qualitative-dominant disclosure this study documents, leading to a failure in assessing the growing environmental footprint by AI infrastructure (Perera et al. 2024: 5).
CONCLUSION
This study examined the extent to which the ten largest Australian financial services companies by market capitalisation disclosed the environmental impacts of AI in their FY2024-25 annual and sustainability reports, and how those disclosures aligned with the requirements of ISSB S2, GRI 302, and GRI 303. The findings reveal a consistent pattern across the sample in which AI is acknowledged as a commercial and operational priority yet systematically withheld from the quantitative environmental accounting that would render its resource consumption transparent. No company in the sample produced AI-disaggregated quantitative energy or water disclosure, with the dominant outcome across disclosure level coding being complete absence or qualitative acknowledgment only. Where environmental disclosure did occur, it gravitated toward organisation-wide aggregates such as Scope 2 carbon emissions and broad environmental impact statements, which satisfy formal framework compliance without requiring AI-specific attribution. The ISSB S2 alignment findings demonstrated a pronounced concentration in governance and strategy pillars alongside a marked deficit in metrics and targets, indicating that companies are comfortable articulating board-level oversight and strategic intent while declining to produce the quantitative performance measures that would allow those declarations to be evaluated. GRI 302 and GRI 303 alignment followed an analogous pattern, with companies citing framework compliance without disaggregating AI or data centre loads within their reported energy and water figures.
Further, our findings are consistent with global literature on AI environmental disclosure rather than as evidence of distinctively Australian underperformance. The disclosure gap documented here reflects structural limitations located at the level of the frameworks themselves, which were designed before AI infrastructure became a material reporting consideration and which remain sector-agnostic in their application. The absence of AI-specific metering standards, the voluntary status of AASB S2 during the study period, and the asymmetric reputational incentives surrounding AI capability disclosure versus AI environmental cost disclosure together account for the thin, qualitative-dominant pattern observed across the six nodes. The findings suggest that voluntary GRI alignment alone is insufficient to surface AI-specific environmental impacts, and that sector-specific guidance within the AASB S2 implementation schedule represents a necessary regulatory intervention if investors and regulators are to obtain the material information required to assess the environmental footprint of a sector whose AI infrastructure dependency is rapidly deepening. Future research should address the limitations of this study by expanding the sample to include smaller financial services entities, undertaking inter-coder reliability testing to strengthen methodological rigour, and conducting longitudinal analysis to track disclosure evolution as mandatory AASB S2 compliance takes effect.
APPENDIX
Figure 1: List of Companies included in study
Comments