A Machine Learning and Econometric Framework for Credibility-Aware AI Adoption Measurement and Macroeconomic Impact Assessment in the Energy Sector

Abstract

Artificial intelligence (AI) adoption in strategic sectors such as energy is often framed in optimistic narratives, yet its actual economic contribution remains under-quantified. This study proposes a novel, multi-stage methodology at the intersection of machine learning, statistics, and big data analytics to bridge this gap. First, we construct a media-derived AI Adoption Score using natural language processing (NLP) techniques, including dictionary-based keyword extraction, sentiment analysis, and zero-shot classification, applied to a large corpus of firm-related news and scientific publications. To enhance reliability, we introduce a Misinformation Bias Score (MBS)—developed via zero-shot classification and named entity recognition—to penalise speculative or biased reporting, yielding a credibility-adjusted adoption metric. Using these scores, we classify firms and apply a Fixed Effects Difference-in-Differences (FE DiD) econometric model to estimate the causal effect of AI adoption on turnover. Finally, we scale firm-level results to the macroeconomic level via a Leontief Input–Output model, quantifying direct, indirect, and induced contributions to GDP and employment. Results show that AI adoption in Romania’s energy sector accounts for up to 42.8% of adopter turnover, contributing 3.54% to national GDP in 2023 and yielding a net employment gain of over 65,000 jobs, despite direct labour displacement. By integrating machine learning-based text analytics, statistical causal inference, and big data-driven macroeconomic modelling, this study delivers a replicable framework for measuring credible AI adoption and its economy-wide impacts, offering valuable insights for policymakers and researchers in digital transformation, energy economics, and sustainable development.

1. Introduction

AI technologies have become increasingly integrated into existing energy systems to significantly improve operational efficiency, enhance smart infrastructure projects, and facilitate a transition to cleaner and sustainable energy systems. AI plays an important role, with applications ranging from predictive maintenance and demand forecasting to intelligent grid management and energy trading to help optimize production, distribution, and consumption. In many emerging economies, including Romania, the actual diffusion and the economic impact of AI on strategic sectors, on the other hand, are underexplored and poorly quantified.

Romania’s energy sector is facing three transformations simultaneously: digitalization, decarbonization, and decentralization. Energy is both a challenge and an opportunity for AI integration as one of the country’s most capital-intensive and technologically rigid industries. Although anecdotal evidence and corporate communications indicate an increasing interest in digital innovation, there is little empirical research on which firms are adopting AI, how credible those claims are, and what economic outcomes follow from this adoption.

In this paper, we fill this gap by offering the first integrated, data-driven assessment of how AI is taking shape in Romania’s energy sector. Through a novel, multi-stage approach, we quantify the AI implementation at the firm level as well as the macroeconomic impact. We combine natural language processing (NLP) techniques to construct a media-based AI Adoption Score, credibility-adjusted through a Misinformation Bias Score (MBS), with econometric modelling and input–output analysis. This framework enables us to step beyond case studies or declarative surveys and provide replicable, scalable insights into the structure, credibility, and impact of AI diffusion.

For the firm-level analysis, we estimated the causal impact of AI adoption on turnover using the Fixed Effects Difference-in-Differences (FE DiD) model. At the macro level, we used these estimates in a Leontief Input–Output model to estimate downstream economic and social impacts, including contributions to GDP and employment. The results show that AI adoption is linked to significant performance and economic output gains, but they also stress the need for targeted public policy to redress uneven adoption, displacement of the workforce, and other challenges around infrastructure.

By connecting media signals with econometric rigour and macroeconomic modelling, this study makes a new contribution to the growing literature on digital transformation in emerging economies. In doing so, it complements existing reports by providing both analytical tools and evidence-driven insights to policymakers, industry leaders, and researchers who seek to understand and encourage the responsible diffusion of AI in critical sectors.

We propose an original methodological and empirical contribution to digital transformation research in emerging markets, combining natural language processing approaches with firm-level econometric modelling and at the macro-level, Input–Output analysis. Unlike existing studies, which mainly depend on declarative questionnaires or aggregated administrative data, we provide a replicable and scalable framework which includes:

• Quantifies an AI Adoption Score based on media-derived signals, offering an indirect yet updatable and comparable perspective across firms;
• Adjusts this score using Misinformation Bias Scores (MBSs), constructed through zero-shot classification and linguistic profiling, to account for speculative or distorted narratives;
• Integrates these estimates into a Fixed Effects Difference-in-Differences (FE DiD) model to derive robust causal effects on firm performance;
• Scales the results to the national level through the Leontief model, estimating the impact of AI on GDP and employment in a context-sensitive and Romania-specific manner;

This approach serves as a new analytical instrument for identifying firms with true AI adoption potential, important for academic research and for formulating public policies accordingly. It also opens novel research opportunities on the interaction of media narratives, technological investment and economic performance in strategic sectors.

The structure of the paper is as follows: Section 2 provides a holistic review of the literature examining the study within the context of ongoing conversations around productivity, organisational change, and the growing threat of AI-generated disinformation. Section 3 describes data and methodology, including how we constructed a firm-level AI Adoption Score based on a dataset of news and scientific content using natural language processing (NLP) techniques, along with the calculation of a Misinformation Bias Score meant to penalise speculative or unverifiable narratives. This dual scoring mechanism feeds into a Fixed Effects Difference-in-Differences (FE DiD) econometric model that estimates the firm-level impact of AI and is further applied to an economy-wide analysis using the Leontief Input–Output Model to estimate the impacts on GDP and employment. Section 4 presents the empirical findings, including comparative AI adoption classifications, estimated treatment effects, and quantified macroeconomic contributions of AI-related activities. Section 5 summarises these results before returning to the implications for credibility, sustainability, and inclusive growth. Section 6 discusses key insights, limitations and policy recommendations, advocating for implementing institutional safeguards that guarantee the responsible deployment of AI in contexts susceptible to information disorder.

2. Literature Review

To build a coherent understanding of AI’s economic impact, this review follows a layered structure. It begins with global implications, narrows down to the European Union’s regulatory and technological landscape, and finally focuses on Romania’s national energy transformation. This cascading perspective helps situate Romania’s experience within broader policy, technological, and institutional dynamics, while also setting the stage for the empirical approach proposed in this study.

2.1. The Global Economic Impact of Artificial Intelligence

Artificial Intelligence (AI) is revolutionizing world economies by automating processes, improving productivity and fostering innovation. While early studies, such as [1], which projected that AI could contribute up to $13 trillion to global output by 2030, more recent research stresses the uneven and disruptive character of this transformation. Dwivedi et al. [2] argue that while AI significantly increases innovation and efficiency, it also deepens existing inequalities and creates systemic vulnerabilities, particularly in labour markets and regulatory frameworks.

In addition to automation and productivity gains, there is an increasing amount of literature concerned with the economic consequences of AI-generated disinformation. The emergence of Generative AI tools, like deepfakes and synthetic text models, creates new vectors of economic distortion. According to Chesney and Citron [3], deepfakes (i.e., hyper-realistic fake videos created using GANs) could threaten financial markets by spreading false information that can manipulate stock prices, disrupt political stability, or damage corporate reputations. This intersection between AI and market perception is particularly relevant in energy sectors, where firms increasingly rely on public ESG narratives to attract green capital. Similarly, recent discussions point out that AI content generation can be used to automate misleading ESG disclosures, generating overly favourable sustainability narratives that do not correspond to actual environmental performance. As stated in the Sustainability Directory [4], AI technologies can exploit data inconsistencies and generate misleading reports that sound convincing but are factually inaccurate, making greenwashing a topic of concern and leading to the potential misallocation of investor capital based on false accusations. These risks are amplified when AI systems generate content without transparent data provenance or human oversight, creating a credibility gap in ESG communications.

Further, Park et al. [5] note that AI-assisted disinformation campaigns corrode institutional trust, which is an essential condition for market stability and investment. One example, with implications in energy and sustainability sectors, is AI-generated greenwashing, where firms exaggerate their environmental performance through false and inflated statements. As Moodaley [6] demonstrated, these practices are deceptive to consumers and investors but damage competitive dynamics and stall meaningful innovation.

2.2. Current Landscape of AI Adoption in the EU

The emergence of artificial intelligence (AI) is revolutionising European economic, social, and industrial systems: there is enormous potential for innovation, but significant regulatory and ethical issues are also involved. AI has been positioned at the EU level as both a strategic driver of productivity and competitiveness, but also as an arena in need of tight governance that embeds responsible, transparent and equitable implementation. This is guided by the European Commission’s AI strategy and the proposed AI Act, which seeks to create common rules, foster trustworthy AI and reduce fragmentation among Member States.

However, AI adoption is still highly uneven within the EU, revealing deep structural differences between countries. At the level of enterprises pursuing innovative expectations, recent Eurostat data [7,8] reveal that in 2023, merely 8% of European enterprises reported using AI technologies. Northern and Western Member States like Denmark (15.2%) and Finland (15.1%) showed the highest levels of adoption, bolstered by developed digital infrastructure, highly effective public–private partnerships and AI-focused innovation policies. On the other hand, there are many countries in Central and Eastern Europe, such as Romania (1.5%), Serbia (1.8%), and Bulgaria (3.6%), that are far behind in this regard, struggling with infrastructure deficits, low levels of digital literacy, and insufficient institutional support [8].

These differences are more than statistical, and they result in imbalances in AI capability, robustness, and information integrity. Countries with low adoption rates may become vulnerable to informational dependencies, where domestic institutions and markets rely on AI-generated insights or technologies developed elsewhere, with limited capacity for local validation. This creates a possible situation of imbalances in information, where less digitally developed economies could be more exposed to AI automatic disinformation, biased automation, or even greenwashing through unverifiable sustainability discourses. In these contexts, generative AI tools, particularly when used for automated reporting, ESG disclosures, or policy simulation, may create greater opacity rather than transparency, especially where local regulatory and auditing frameworks are weak.

The European Union has positioned itself as a leader in the ethical governance of AI. In this regard, the European Strategy on Artificial Intelligence and the Artificial Intelligence Act aim to harmonise AI regulation in various Member States when it comes to transparency, accountability, and risk management. Floridi et al. [9] highlight the need to align AI adoption with ethical values to mitigate risks and ensure public trust. Furthermore, as pointed out by Winfield and Jirotka [10], there is an urgent need for regulatory frameworks that balance innovation while considering human rights.

2.3. Adopted AI Technologies and Practices Among European Firms

The technologies that European enterprises adopted most widely included text-mining systems, speech recognition tools, and machine-learning paradigms. They can facilitate data analysis, predictive decision-making and process optimisation. In this way, machine learning (ML) and deep learning (DL) are used to achieve similar predictive capable engines in application domains such as finance, healthcare, and manufacturing [11]. These technologies underscore the importance of AI in driving competitiveness and innovation across European sectors.

Recent research indicates the central role that ML and DL are starting to play in transforming Europe’s energy landscape. AI techniques, especially deep learning and neural networks, have proved beneficial for short-term energy forecasting, a key element to achieving grid stability and smart energy. These approaches assist in the development of smart grids and allow the establishment of collaborative energy-sharing systems, which represent one of the major challenges on the road to clean energy in Europe [12].

In renewable energy infrastructure, AI systems play an important role—with applications ranging from optimising the deployment of solar and wind resources to detecting equipment faults and automating the operation of energy consumption in smart cities. In such applications, AI-enabled technologies contribute to achieving sustainability targets while improving productivity and economic resilience for the energy industry [13].

These insights reinforce the growing relevance of AI technologies beyond traditional business operations, particularly in enabling Europe’s transition toward clean and intelligent energy systems.

2.4. AI as a Strategic Advantage in the Organisational Environment

Artificial intelligence (AI) application is a significant competitive advantage for the enterprise. With AI-powered applications, companies can enhance their operations, boost decision-making skills, and catalyse innovation. On the other hand, AI’s inclusion in the business environment brings about obstacles, such as the cost and the inability to incorporate AI tools into the workforce.

There are several operational and strategic advantages of integrating AI into organisational processes. Automation, predictive analytics, and complex problem-solving that AI can automate, improve and further enhance efficiency while reducing costs. Cannas et al. [14] explain how AI can have a tremendous impact by revolutionising the supply chain and operations management as AI systems reduce the lead time for inventory management, enhance service quality and allow firms to minimise wastage of their resources. Likewise, Al-Surmi et al. [15] point out that AI is being implemented in marketing and information technology to help foster business performance and quality of decision-making.

AI technologies also enhance innovation potential through the introduction of new products and services. For example, Huang and Rust [16] highlight the potential of AI in customer engagement in the service sectors, noting its ability to relieve customer service jobs from repetitive tasks while amplifying the quality of engagement with the customer by customising their experience.

2.5. Organisational Challenges in AI Implementation

Although AI technology could yield considerable benefits, there are some barriers to its successful adoption across organisations. According to Ipsos [17], European companies face many challenges related to AI implementation, such as high costs (52%), difficulties in adapting existing operational processes (49%) and a shortage of people with sufficient qualifications (57%). In addition, Tariq [18] states that a lack of IT infrastructure, poor quality data and regulatory uncertainties are also obstacles to the use of AI. These problems are compounded by resistance to change in organisational cultures and a lack of strategic planning for AI integration. Operational inefficiencies could also stem from weak data governance and an opaque AI system. Rana et al. [19] discuss how suboptimal decisions stemming from unreliable AI-based analytics can create risks and erode competitive advantage. These barriers to implementation can be particularly acute in strategic sectors like energy, where real-time decisions depend on the accuracy and transparency of AI-driven forecasts and control systems.

AI is strategically important because it can enable digitalisation and create new streams of business value. Enholm et al. [20] describe that the foundation for strategic positioning in a competitive market is given through AI assisting organisational management, replacing ineffective tasking as well as motivating organisational innovations. Similarly, Kitsios and Kamariotou [21] highlight the significance of artificial intelligence in aligning IT with organisational objectives to enable knowledge management to innovate and update services, leading to the development of new sources of revenue. Ahmad et al. [22] provide evidence from multiple energy markets showing that AI adoption significantly improves grid performance and load balancing. Their work demonstrates how deep learning and hybrid models enhance demand forecasting, fault detection, and generation scheduling—contributing directly to cost efficiency and emissions reduction. Integrating such sector-specific use cases into organisational AI strategies remains a key opportunity, especially in countries seeking to modernise outdated infrastructure while meeting climate targets. A strategic approach is the best way for organisations to leverage the gains of AI while minimising the risks. So, the public and private sectors, academia, and policymakers should work together to create best practices and close knowledge gaps. This would ensure an environment that enables responsible AI adoption and compliance with ethical standards.

2.6. Economic Analysis Models in the Context of AI

As AI adoption has rapidly increased across industries, the economic effects of this growth need to be assessed through national productivity, in addition to firm performance scalability models. A new area of research relies on quantitative approaches to appreciate the impact of AI on firm performance, persistence, and systemic change.

Alekseeva et al. [23] make a foundational contribution using firm-level data and a panel regression framework based on firm-specific AI-related job postings to evaluate the impact of AI adoption on firm performance. Their empirical results show a significant positive relationship between AI adoption and growth in firm size, capital investment, and R&D spending. Notably, the performance advantages were largely attributable to AI-related expertise among the management team, not to IT staff. However, no significant effects were found on productivity metrics, indicating that AI’s contribution may lie more in expansion and innovation than in operational efficiency.

Chen and Tajdini [24] also studied the phenomenon in a complementary paper, implementing a moderated regression model to assess organisational, technological, and environmental factors affecting AI adoption intensity. Their findings, which come from firm survey data, indicate that higher AI adoption is positively correlated with firm performance, especially under conditions of technological turbulence. This means that external and contextual variables are key in supporting the benefits that firms can extract from AI.

Examining AI’s roles in instances of systemic stress, Ho et al. [25] used Interrupted Time-Series Analysis (ITSA) to analyse firm resilience during the COVID-19 pandemic. The analysis indicated that firms possessing pre-existing AI capabilities faced less severe negative financial impacts and were quicker to recover their stock performance than their less digitally enabled peers, suggesting that AI can serve an economic shock-doctor role against current events and market volatility.

Kim, Park, and Kim [26] used a Difference-in-Differences (DiD) analysis of a panel of 105 U.S. firms from 2008 to 2014 to identify causal effects. Their analysis differentiates between the two types of AI, automation-related vs. augmentation-related, and finds that only the former significantly improved the value and cost efficiency of the firm. Focusing on within-industry effects, this approach increases the robustness of causal inferences and provides a nuanced understanding of the heterogeneous effects of different AI applications.

Although these micro-level studies give important evidence of how AI influences firm dynamics, insights into the macroeconomic and environmental impacts of AI adoption on a systemic level necessitate more systemic analytical tools. This task suits analysis through Input–Output (IO) modelling, especially via the Leontief framework. It captures the intersectoral dependencies and allows direct, indirect, and induced effects to be quantified throughout the economy.

In a recent paper [27], the authors demonstrate the applicability of this methodology in emerging economies by using the Leontief matrix in order to measure the impact exerted by diverse categories of investment (namely, physical capital, R&D, and human capital) on Romania’s global output. They found that the volatile coefficients of input–output interdependencies in transitional economies entail the need for sectoral strategies adapted to dynamic structures and technologies. Likewise, Davidescu et al. [28] used a green Leontief IO model to evaluate the environmental ramifications of EU Structural and Investment Funds, showcasing the strength of the methodology as it decomposes direct, indirect and induced impacts from policy interventions. Although their study specifically targets green finance, the same modelling approach is applicable for estimating the broader macroeconomic implications of AI adoption, including spillover effects to GDP and employment, through interconnected sectors.

In summary, a combination of firm-level causal approaches (such as DiD and ITSA) with macroeconomic IO modelling provides a solid analytical set of tools to assess AI’s dual impact on productivity and sustainability across the micro and macroeconomic fronts.

2.7. Artificial Intelligence in the Context of Energy and Climate Challenges

AI’s uses within the energy industry are promising for the advancement of innovation and economic development. One of the latest reports outlines how the use of AI predictive maintenance systems for power plants can help reduce equipment downtime, which would further embed energy generation facilities into a more sustainable operations paradigm [29]. Moreover, AI tools are crucial to demand-side management strategies, with the ability to predict consumption patterns, optimize energy allocation, and reduce peak load stress on grids. Rolnick et al. [30] refer to it as an element that, if developed, can be used to solve climate change issues by organising resources and reinforcing climate modelling. Similarly, according to Rockström et al. [31], technologies must work in harmony with the planetary boundaries required for sustainability.

Vinuesa et al. [32] emphasize the dual role of AI in accelerating progress toward the Sustainable Development Goals (SDGs) while also posing new ethical and environmental risks. Their analysis identifies energy systems as a critical domain where AI can enhance efficiency through real-time optimization, predictive maintenance, and renewable energy forecasting. At the same time, they caution that without proper governance, AI may exacerbate existing inequalities and reinforce unsustainable practices—particularly when models are trained on biased or incomplete data. These insights underscore the need to embed transparency, auditability, and credibility-aware metrics into AI systems deployed in climate-critical sectors.

However, misinformation and manipulation are rising as AI, especially generative systems like Large Language Models (LLMs), becomes more embedded into sustainability reporting and communication. De Villiers et al. [33] caution that generative AI can be used to produce persuasive yet unverifiable sustainability content, simulating environmental progress through language that may not reflect actual performance. They highlight how tools like ChatGPT-4o could be used to automate ESG disclosures that exaggerate achievements, reference unsupported certifications, or omit negative outcomes, potentially giving stakeholders the wrong impression. Senni et al. [34] show that generative NLP tools applied to corporate transition plans tend to express high-level target setting (“talk”) but reveal substantially fewer concrete implementation strategies (“walk”). Their findings, based on the analysis of sustainability disclosures from Climate Action 100+ companies using an LLM-based tool, suggest that this discrepancy may be signalling greenwashing behaviour, where vague but optimistic language masks a lack of verifiable action. Companies tend to publish more about goals than about tangible, operational changes, which complicates efforts to assess true climate commitment.

2.8. Romania’s Energy Sector Transformation and the Contribution of AI

The Romanian energy landscape has been progressively shaped by the transition from fossil fuel dependency to the increasing use of green energy and is undergoing significant transformations driven by Artificial Intelligence (AI). Notably, as the nation embraces the European Green Deal, AI is critical for integrating renewable energy sources, optimising grid management, and promoting sustainability.

Romania was an important source of oil in the nineteenth and twenty-first century, with Ploiești emerging as the heart of oil extraction and refinement. This took Romania to the fore of the first countries to export petroleum abroad. Historical milestones such as these are well-documented by Tulucan et al. [35], which highlights Romania’s pioneering role in petroleum refinement and export.

Once communism took hold in Romania, the growing use of coal and hydroelectric power was enough to fuel the rapid growth in industrialisation. The partnership with the Yugoslavians on large-scale hydro construction projects, such as the Iron Gates showcased Romania’s commitment to energy independence. This period and its achievements are analysed by Rugina [36], which outlines Romania’s strategic energy collaborations and infrastructure projects.

As the Romanian market gradually liberalised and the country became a member of the EU, the planning policies changed with a focus on the development and more efficient usage of energy sources. Romania shows excellent results in this regard as the renewables share in the gross final energy consumption reached 24% in 2020, much above the EU goals. These developments are discussed by Colesca and Ciocoiu [37], who provide an in-depth review of Romania’s renewable energy policies and achievements.

2.9. AI Technologies and Advances in Romania’s Energy System

AI is key to optimising renewable energy systems in Romania, especially for wind and solar energy. Murărașu [38] observes the volatility of renewable energy production and mentions the role of AI-driven models in stabilising energy outputs and enhancing the economic integration of renewable sources. The study demonstrates that these AI techniques significantly reduce balancing costs, improve forecasting accuracy, and contribute to Romania’s transition to a low-carbon economy.

Another study [39] promotes Romania’s projects for increasing wind and solar production through AI-optimised grid systems. The research demonstrates how AI technologies can boost reliable seasonal energy production, aiding Romania in realising its renewable energy objectives. Technologies for smart grids powered by AI have been essential in updating Romania’s energy systems. Crețulescu and Crețulescu [40] show that machine learning techniques such as XGBoost can effectively predict energy consumption and alleviate price volatility, which ultimately aids in stabilising the national grid. The study also illustrates how AI can address infrastructure issues, such as the temporary shutdown of Cernavodă Nuclear Power Plant Reactor 1.

Moreover, Frătiţa et al. [41] examine the adaptability of Romania’s grid for renewable energy, emphasising key sites for green energy installations. AI models have played a crucial role in harmonising supply and demand, simultaneously decreasing reliance on energy imports.

2.10. Challenges and Recommendations for AI Adoption in Romania

Although it has potential, AI adoption in Romania faces many significant challenges. Murărașu [38] indicates that the instability of renewable energy sources is one of the main barriers. Their analysis detects anomalies in renewable energy generation, which can only be adjusted with advanced AI systems.

Governance and regulation barriers are still major challenges. In a study of the financial performance of some Romanian energy companies [42], the authors showed lagging AI adoption due to poor policies and weak governance structures. The research also recommends improved governance systems to deal with the widespread absence of AI use in the energy sector. The study concludes that this would promote AI-powered innovation in the energy sector.

Strategic interventions are essential to mitigate these challenges. Cristea et al. [43] recommend expanding financial incentives for AI-driven renewable energy projects. Their study evaluates home photovoltaic systems in Cluj-Napoca and shows that AI-enhanced systems are economically viable and sustainable for residential use.

Furthermore, education in the workforce is the best path to resolving the current skills gap. According to Tănasie et al. [44], there is an increasing need for green jobs in the developing field of renewable energy in Romania, and they advocate specific educational schemes that raise AI skills in the energy industry.

In conclusion, despite the promise of AI in advancing sustainability and energy planning, the recent literature warns that generative models may also amplify informational opacity, creating a new form of algorithmic disinformation masked under the guise of transparency. This is especially concerning in domains like ESG reporting, where unverifiable claims produced by LLMs can mislead investors and distort public accountability.

These challenges are aggravating in emerging economies, where structural weaknesses in regulatory oversight, digital infrastructure, and institutional capacity may constrain socioeconomic validation of AI-generated insights. In these contexts, such as in the Romanian case, generative AI tools may entrench existing asymmetries, not only in innovation but also in information control and governance.

These intertwined dynamics of economic transformation and informational distortion highlight the urgency of integrating credibility-aware metrics in AI research, a core concern of the ongoing scholarly dialogue on generative AI and information disorder.

Thus, this study constitutes a substantial contribution to the literature, developing an original firm-level AI Adoption Score adjusted via a Misinformation Bias Score (MBS) to account for inflated or unverifiable claims, an issue increasingly relevant in the age of generative AI. Focusing on Romania’s energy sector, we use a Difference-in-Differences (DiD) model to assess the financial impact of AI at the individual firm level and combine it with a Leontief Input–Output model to evaluate economy-wide effects. By bridging microeconomic causality with macroeconomic architecture, our framework provides a comprehensive and credibility-sensitive measure of AI’s transformative power. In doing so, we address recent calls in the literature, particularly from Vinuesa et al. [32] and De Villiers et al. [33], to integrate qualitative and credibility-aware indicators into quantitative models of AI impact.

3. Materials and Methods

The methodology is structured in multiple stages: it combines AI Adoption Scoring, econometric modelling, and macroeconomic impact assessment. The AI Adoption Score is calculated using NLP techniques applied to media articles to provide a proxy for firm-level engagement with AI technologies. But while this score gives a sense of the visibility and intensity of AI-related activity, it does not necessarily take into account the credibility or fact-based reliability of the narratives from which it is assembled. In some cases, particularly where AI is linked to innovation and sustainability, companies and stakeholders may engage in discursive amplification, using vague, speculative, or overly optimistic language to portray initiatives as more advanced than they truly are. This type of distortion, often called “AI greenwashing”, can mislead public perception and policy, leading to misdirected incentives, exaggerated economic projections, or decisions based on perceived rather than actual innovation.

To mitigate this risk, we propose a Misinformation Bias Score (MBS) that captures exaggeration and speculative framing in AI-related media discourse. This allows us to develop a Penalised AI Adoption Score, which is adjusted for the quality of information. Combining the raw and penalisation scores, we determine two groups of companies: the treatment group, consisting of firms that have actively integrated AI into their business models, and the control group, composed of firms that continue to follow traditional operational practices without evident AI engagement. These classifications are used in the Fixed Effects Difference-in-Differences (FE DiD) estimation to evaluate the causal impacts of AI adoption on firm performance. The consistency of the treatment effect across both scoring methods confirms the robustness of our classification approach. Moreover, the estimated impact is applied to reach the share of turnover or value-added associated with AI activities, which is included in the Leontief Input–Output Model to determine the direct, indirect, and induced AI adoption contributions on GDP and employment, providing a comprehensive and credibility-adjusted reflection of AI’s macroeconomic impact.

3.1. Data Collection

The primary data source for this study is the National Trade Register Office (NTRO) database, which includes over 676,000 registered companies in Romania. Focusing on the energy sector (Appendix A.1), we selected firms classified under NACE code 35 (energy production and distribution), narrowing further to two key subsectors: electricity production (3511) and electricity trading (3514). These were chosen for their central role in the energy market and high potential for AI integration in production, storage, and distribution. From the 2483 firms in this category, we selected a sample of 36 companies, representing 87% of the sector’s total turnover. Selection criteria included financial strength, market influence, and AI adoption potential, ensuring the sample reflects the sector’s economic core and broader industry trends. To assess AI adoption at the company level, we developed an AI Adoption Score built from unstructured data sources such as news articles and scientific publications. We also investigated the availability of more data sources—such as AI-related patents, formal job postings, and corporate innovation disclosures—but found these signals to be largely absent in the context of Romania’s energy sector. Most firms are operational rather than research-oriented, and relevant documents rarely reference AI in a systematic or verifiable manner. Using natural language processing (NLP) techniques, including sentiment analysis and keyword extraction, we processed a dataset of 500 news articles and 2399 AI-related scientific abstracts to capture relevant adoption signals. These insights supported the creation of a robust AI dictionary and scoring mechanism, detailed further in Appendix A.2.

3.2. AI Adoption Scoring Development

The AI Adoption Score quantifies how actively companies in Romania’s energy sector are embracing artificial intelligence, serving as a key input for assessing AI’s economic relevance. To construct this score, we applied natural language processing (NLP) techniques to a corpus of media content and scientific literature, enabling the extraction of AI-related signals at the firm level. The process integrates dictionary-based keyword analysis, sentiment evaluation, and score normalisation to produce a comparable adoption metric across companies. The scoring methodology followed several key steps. First, a specialised AI dictionary was constructed from high-impact academic abstracts extracted from the Web of Science database. Terms were lemmatised and validated using GPT-4 Turbo to ensure relevance and clarity. Each keyword was then scored using the AFINN sentiment framework, assigning weights between 1 and 5 based on their tone and context, and visualised through a word cloud to highlight dominant AI themes. In parallel, the collected news articles were pre-processed through tokenisation, lemmatisation, and noise removal to isolate meaningful AI-related content. Each article was then evaluated for sentiment using AFINN, and average sentiment scores were calculated per company, forming the initial AI Adoption Score. These scores were subsequently normalised on a scale from 1 to 10 using the Min-Max method to enable cross-firm comparison. Finally, based on score thresholds (mean, median, and third quartile), companies were classified into treatment and control groups, identifying those more advanced in AI integration versus those maintaining traditional practices. This classification informed the econometric design of the Difference-in-Differences (DiD) analysis conducted in the later stages of the study.

Enhancing Source Credibility in AI Adoption Scoring: Misinformation Bias Score (MBS)

To improve the validity of the AI Adoption Score derived from media sources, we introduce a Misinformation Bias Score (MBS) designed to apply a penalty based on the linguistic quality of an article. This adjustment is implemented via two complementary methods: (1) zero-shot classification at the article level and (2) linguistic profiling using Named Entity Recognition (NER).

For the article-level classification, the limited corpus size (~500 articles) didn’t facilitate the fine-tuning of a domain-specific model without a significant risk of overfitting. Consequently, we adopted a zero-shot classification approach, selecting the knowledgator/gliclass-modern-large-v2.0 model [45]. This choice was informed by its state-of-the-art performance and its specific suitability for analyzing journalistic text. For instance, on the AG_NEWS news classification benchmark, this model achieves a Macro F1 score of 0.736, demonstrating superior performance compared to other widely used zero-shot models, such as Moritz/deberta-v3-large-mnli-fever-anli (F1 score: 0.612). Furthermore, its ModernBERT architecture, extended context capacity (4096 tokens), and faster inference times make it suitable for processing full-length news articles.

Using this model, each article is classified as either “Objective Reporting” or “Potentially Biased or Misleading,” with the probability score for the latter category serving as the heuristic penalty for the MBS. While we acknowledge that any general-purpose model has inherent limitations in capturing all domain-specific nuances—a challenge amplified by the lack of a large, annotated corpus for validation—the chosen model’s empirically demonstrated robustness on diverse news classification tasks makes it a highly suitable and reliable proxy for source quality. The MBS should therefore be interpreted as a data-driven heuristic to moderate the AI Adoption Score, rather than a definitive measure of factual accuracy.

In our second method, we perform a granular linguistic profiling of each article to augment the classification score. This task required the identification and quantification of nuanced linguistic categories that signal objectivity or bias—categories for which no standard, pre-trained Named Entity Recognition (NER) models exist. Traditional NER systems are limited to a fixed set of entities (e.g., persons, locations) and are unsuitable for this type of semantic analysis.

To overcome this limitation, we employed urchade/gliner_large-v2 [46], a state-of-the-art model specifically designed for flexible, zero-shot NER. Its key architectural advantage is the ability to recognize custom entity types defined on-the-fly, making it uniquely suited for our research needs. We configured the model to extract the following six categories, which are indicative of an article’s framing and degree of factuality:

• Quantitative data (e.g., statistics)
• Attribution or source (e.g., “according to the report,” “the CEO stated”)
• Hedging language (e.g., “could,” “might,” “potentially”)
• Subjective opinion (e.g., “a remarkable achievement,” “disappointing results”)
• Emphasis or exaggeration (e.g., “completely,” “never,” “huge”)
• Forward-looking statements (e.g., “will launch,” “plans to develop”)

For each article, we compute the relative frequency of these entities against the total word count, yielding a quantitative measure of its factuality and speculative tone. To ensure comprehensive analysis within the model’s context limits, articles exceeding a certain token length were segmented by sentence, with entity frequencies aggregated across all segments. This approach provides a data-driven profile of the linguistic characteristics of each source.

3.3. Integration of the MBSs into the Adjusted AI Adoption Score

Both Misinformation Bias Scores (MBSs) have been normalized to the [0, 1] interval and integrated into the AI Adoption Score adjustment formula as follows:

$${\mathrm{A}\mathrm{I}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_{\mathrm{p}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{z}\mathrm{e}\mathrm{d}}=\mathrm{A}{\mathrm{I}}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}{\mathrm{e}}_{\mathrm{r}\mathrm{a}\mathrm{w}}}\times \left(1-\alpha \times \mathrm{M}\mathrm{B}{\mathrm{S}}_{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{d}}\right),$$

where ${\mathrm{A}\mathrm{I}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_{\mathrm{p}\mathrm{e}\mathrm{n}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{z}\mathrm{e}\mathrm{d}}$ is the final, credibility-adjusted score; $\mathrm{A}\mathrm{I}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}{\mathrm{e}}_{\mathrm{r}\mathrm{a}\mathrm{w}}$ is the initial AI Adoption Score computed prior to adjustment; α is an empirically determined penalty parameter and $\mathrm{M}\mathrm{B}{\mathrm{S}}_{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{d}}$ is the Misinformation Bias Score.

In this study, the Misinformation Bias Score (MBS) is computed using the two complementary approaches previously described: (1) the article-level zero-shot classification score (Z-bias score) and (2) the linguistic profiling score via zero-shot NER (NER-bias score). A penalty coefficient, α, is applied to moderate the final AI Adoption Score based on the MBS.

The selection of this coefficient was empirically guided by the intrinsic structure of our data. To determine an appropriate value, we first performed an unsupervised K-Means clustering analysis on the articles. The stability of the resulting data partition was subsequently confirmed using the HDBSCAN algorithm, with an overlap analysis indicating that both methods identified robust and highly similar clusters. This analysis revealed a natural partition of the data into two distinct clusters: a large cluster (72.1% of articles) characterized by high objectivity and a smaller, well-defined cluster (27.9%) characterized by potentially biased or speculative narratives. Crucially, the mean objectivity score for this second cluster was 0.31. This finding provided a strong, data-driven anchor for our coefficient, leading to the selection of α = 0.3.

To validate the robustness of this empirically derived value, we subsequently conducted a sensitivity analysis by varying α within the range of [0.1, 0.5]. This analysis confirmed that α = 0.3 provides the most balanced adjustment, meaningfully penalizing biased narratives while preserving the stability of the treatment group classification and the consistency of the Fixed Effects Difference-in-Differences (FE DiD) model estimates.

Thus, companies with a high AI Adoption Score but associated with speculative or biased media coverage received a lower adjusted score. In contrast, those with more objective and factual media coverage maintained their original score.

This adjustment allows for a more prudent and context-aware measure of AI adoption, amplifying the robustness of econometric and macroeconomic implications developed through subsequent iterations of the research. From the scores without and with AI adoption penalty provided, we distinguish the main groups of firms required for the impact evaluation phase: one treatment group, which includes firms that have actively adopted AI in their business models and one control group, which includes companies that still follow traditional operational routines without apparent engagement in any AI related activity.

3.4. Estimating the Share of Turnover Linked to AI-Related Activities in the Energy Sector: A FE DiD Approach

To identify the causal effect of AI adoption on firm performance in Romania’s energy sector, we apply a Fixed Effects Difference-in-Differences (FE DiD) model, as seen in Figure 1. This method controls for unobserved, time-invariant characteristics and isolates the impact of AI adoption on financial outcomes—specifically, turnover and net profit.

Companies are classified into treatment and control groups based on both the AI Adoption Score and the Penalized AI Adoption Score (adjusted using the Misinformation Bias Score). Firms with scores above selected thresholds (mean, median, third quartile) form the treatment group (AI adopters), while those below form the control group (non-adopters).

Firm-level data for 2019 (pre-adoption) and 2023 (post-adoption) was retrieved from the National Trade Register Office (NTRO). Outcome variables—turnover and net profit—were adjusted for inflation (CPI, 2019 = 100). Control variables include firm size (log of employees), own capital (log), fixed assets, and firm age. The FE DiD model is specified as follows:

$${\mathrm{l}\mathrm{n}(\mathrm{Y}}_{\mathrm{i}\mathrm{t}})={\alpha }_{\mathrm{i}}+{\delta }_{\mathrm{t}}+{\beta }_1{\mathrm{D}}_{\mathrm{i}}+{\beta }_2{\mathrm{T}}_{\mathrm{t}}+{\beta }_3{(\mathrm{D}}_{\mathrm{i}}\ast {\mathrm{T}}_{\mathrm{t}})+{\gamma }_{\mathrm{i}}{\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{ϵ}}_{\mathrm{i}\mathrm{t}},$$

where

• ln(Yit): Natural logarithm of the outcome variable (turnover or net profit) for firm i at time t, adjusted for inflation.
• ${\alpha }_{\mathrm{i}}$: Firm fixed effects capturing unobserved, time-invariant characteristics specific to each firm.
• ${\delta }_{\mathrm{t}}$: Time fixed effects accounting for macroeconomic shocks or time-specific factors affecting all firms.
• ${\mathrm{D}}_{\mathrm{i}}$: Binary indicator equal to 1 if the firm belongs to the treatment group (i.e., adopted AI), 0 otherwise.
• ${\mathrm{T}}_{\mathrm{t}}$: Binary indicator equal to 1 for the post-adoption period (2023), 0 for the pre-adoption period (2019).
• ${(\mathrm{D}}_{\mathrm{i}}\ast {\mathrm{T}}_{\mathrm{t}}$): Interaction term; its coefficient ${\beta }_3$ is the DiD estimator, capturing the causal effect of AI adoption.
• ${\mathrm{X}}_{\mathrm{i}\mathrm{t}}$: Vector of time-varying firm-level control variables including: firm size (log of number of employees), own capital (log), fixed assets, firm age.
• ${\mathsf{ϵ}}_{\mathrm{i}\mathrm{t}}$: Idiosyncratic error term.

We verified the parallel trends assumption by comparing pre-adoption trajectories between treated and control firms. Robust standard errors are used to correct for heteroskedasticity, and model fit is assessed using within R². The coefficient β3 measures the causal change in performance due to AI. A positive β3 indicates a beneficial impact on turnover or profit. To estimate the share of economic activity linked to AI, we use:

$$\mathrm{A}{\mathrm{I}}_{\mathrm{s}\mathrm{h}\mathrm{a}\mathrm{r}\mathrm{e}}=\mathrm{e}\mathrm{x}\mathrm{p} \mathrm{e}\mathrm{x}\mathrm{p}\left({\beta }_3\right)-1$$

Multiplying this share by post-treatment turnover or profit yields the total value of AI-related activities, which is then integrated into the input–output model to assess AI’s contribution to GDP and employment.

3.5. Economic and Social Impact Assessment Using the Leontief Input–Output Model

To assess the broader economic effects of AI adoption in Romania’s energy sector, the total value of AI-related activities (as determined in the previous FE DiD stage) was integrated into a Leontief Input–Output (IO) model. This model evaluates how output from one sector affects production across the economy, quantifying direct, indirect, and induced contributions to GDP and employment. The analysis uses the 2021 symmetric input–output table (64 sectors, basic prices) published by the National Institute of Statistics (NIS). The electricity, gas, steam, and air conditioning supply sector is used as the target subsector. The model focuses on backward linkages, which show how increased demand in the energy sector stimulates upstream production across the economy. We apply both Type I and Type II multipliers: Type I multipliers account for direct and indirect effects, excluding households, while Type II multipliers include induced effects via household income and consumption. The Leontief inverse matrix is computed from the technical coefficient matrix A, which captures sectoral input requirements, and from this there have been used: output multipliers, total output needed across all sectors to support one unit increase in energy sector output, GDP multipliers, GDP generated per unit of sectoral output and employment multipliers, number of jobs created per unit of output, with and without induced effects.

The direct, indirect, induced and total effect of AI adoption on GDP can be determined as follows:

• Direct contribution of AI adoption in the energy sector to GDP formation = Sector Output from AI-related Activities/Total GDP
• Direct and indirect contribution of AI adoption in the energy sector to GDP formation = Sector Output from AI-related Activities × Type I GDP Multiplier/Total GDP
• Direct, indirect, and induced contribution of AI adoption in the energy sector to GDP formation = Sector Output from AI-related Activities × Type II GDP Multiplier/Total GDP

In a similar way, the direct, indirect, induced and total effect of AI adoption on employment (thousands of people) can be determined as follows:

• Direct contribution of AI adoption in the energy sector to employment (number of employed persons) = AI-Related Sector Output × Employment sector/Output sector
• Direct and indirect contribution of AI adoption in the energy sector to employment (number of employed persons) = Type I Employment Multiplier × AI-Related Sector Output × Employment sector/Output sector.
• Direct, indirect, and induced contribution of AI adoption in the energy sector to employment (number of employed persons) = Type II Employment Multiplier × AI-Related Sector Output × Employment sector/Output sector.

4. Empirical Results

This section details the results of each analytical stage—from building the AI dictionary to continuing with insights from the media corpus and then to the firm-level econometric analysis and the macroeconomic assessment via the Leontief input–output model.

As a precursor to those empirical inquiries, a contextual overview of Romania’s energy sector (Appendix A.3) establishes a necessary basis for what follows. The industry’s changing production mix, shifting import-export dynamics, and persistent emissions challenges all highlight the need for technological modernisation and provide a ripe opportunity for advanced digital tools like artificial intelligence. As Romania seeks to align with EU energy and climate goals, the integration of AI becomes a strategic lever for enhancing grid efficiency, forecasting, and sustainability performance. However, this transformation does not take place in a vacuum. Romania’s emerging economy profile, characterised by uneven AI readiness, restrictive regulatory oversight, and gaps in the digital infrastructure, renders it particularly vulnerable to informational vulnerabilities. In these contexts, using generative AI to automate ESG claims or exaggerate innovation narratives can obscure real progress, amplifying the risk of greenwashing and compromising policy credibility. Using a novel AI Adoption Score corrected by a Misinformation Bias Score, our empirical analysis addresses these challenges directly, providing a credibility-aware assessment of digital transition for a critical infrastructure sector.

4.1. Building the AI Dictionary: Identifying Core Concepts of AI Adoption

The AI-specific dictionary, constructed from 2399 highly cited scientific abstracts, reveals key dimensions of AI development and application. The word cloud (Figure 2) highlights the dominance of terms such as “artificial”, “intelligence”, “learning”, and “model”, reflecting the foundational role of machine learning and data-driven approaches in current AI research.

Other frequently occurring terms, including “performance”, “application”, “predict”, and “evaluation”, emphasize the growing focus on measurable impact and utility—particularly relevant for understanding how firms may position AI as a tool for operational optimization or innovation.

This dictionary serves not only as a basis for identifying AI-relevant terms in news content but also offers a conceptual framework to interpret firms’ AI communication strategies in the Romanian energy sector.

4.2. Textual Analysis of News Content: Sectoral Framing of AI and Energy Development

The exploratory analysis of 500 news articles referencing the selected energy firms reveals thematic convergence between innovation, investment, and sustainable growth (Figure 3). Terms such as “project”, “development”, “renewable energy”, and “power plant” confirm the dual positioning of the sector, grounded in traditional energy but increasingly committed to green transformation.

Corporate framing emphasises investment scale and market orientation, as shown by the prevalence of terms like “euro”, “million”, “share”, and “customer”. This aligns with strategic narratives meant to appeal to both investors and regulators.

The language is notably forward-looking, with terms like “growth”, “installation”, “new capacity”, and “photovoltaic park”, suggesting active modernisation. These results validate the inclusion of AI adoption in broader digitalisation and sustainability efforts and justify the next step of measuring how such narratives translate into performance.

4.3. AI Adoption Scores: Raw Versus Penalised Results

The classification of companies by their engagement with AI technologies was performed using two versions of the AI adoption score: a raw score computed based on sentiment and keyword signals derived from media coverage and two penalized scores adjusted for possible narrative distortion using both alternative Misinformation Bias Scores (MBSs) calculated with article-level zero-shot classification, and a linguistic profiling approach based on Named Entity Recognition (NER). Different ways of calculating MBSs have been used: zero-shot classification (classifying articles as objective or biased) and NER-based linguistic profiling (counting hedging, exaggeration and factuality markers).

The comparative bar charts show the classification of Romanian energy companies by the likelihood that a company will adopt AI using three scoring systems: the raw AI adoption score, the penalised AI score (from zero-shot classification) and the penalised AI adoption score (generated from NER profiling). The firms are classified into binary categories, AI adopters (1) or non-adopters (0), and these classifications are calculated under three statistical thresholds: mean, median, and third quartile (Q3).

When the mean is used as the threshold (Figure 4), there is consistency in how the three scoring approaches classify companies across most companies. But Enel Energie Muntenia SA is an exception. Even though it is categorised as a non-adopter based on the raw score (solid colour), both penalised scores (zero bias score and NER bias score) classify it as an adopter (hatched bars), indicating that after adjusting for disinformation or exaggeration, its AI-related discourse seems to be more in line with adoption narratives. The consistency between scores lends confidence to further analyses conducted relying specifically on the Fixed Effects Difference-in-Differences (FE DiD) model, which assumes fixed treatment and control groups over time.

The median-based classification only has minor deviations (Figure 5). The raw score identifies COMPANIA MUNICIPALĂ TERMOENERGETICA BUCUREŞTI S.A. as a non-adopter, but both penalised scores reclassify it as an adopter. Likewise, ENGIE ROMANIA S.A. and RENOVATIO TRADING SRL fall into the category of AI adopters according to the penalised score generated based on zero-shot classification, suggesting that the assessment of credibility can change the classification outcome based on AI.

Classifications based on the third quartile are consistent across all three scoring systems (Figure 6). It implies that a more conservative threshold (Q3) cancels out the differences between the raw and the penalised methodologies, indicating higher agreement when harsher AI adoption criteria are applied.

This shows the value of using AI adoption scoring paired with disinformation-sensitive methodologies. They also demonstrate how methodological choices, particularly related to handling exaggerated or unverifiable statements, can affect the classification of firms. This highlights the need for credibility-aware measures of AI adoption, especially in contexts where greenwashing and inflated digital narratives may obscure actual technological implementation.

4.4. Causal Estimation of AI Adoption Effects on Firm Performance: A Fixed Effects Difference-in-Differences Analysis

This section, building on the classification of firms into treatment and control groups according to their raw and penalized AI Adoption Scores, estimates the causal effect of AI adoption on firm turnover by applying a Fixed Effects Difference-in-Difference model. For treated (AI adopters), the comparison is against control (non-adopters) firms using pre-(2019) and post-treatment (2023) data, using different thresholds (mean, median, followed by third quartile) to define treatment in the model.

Before running the FE DiD model, the parallel trends assumption was verified to ensure that treated and control groups had similar trends in turnover before treatment. This supports the validity of including only the interaction term and covariates, as the parallel trends assumption justifies that unobserved differences between groups are time-invariant and absorbed by the fixed effects.

The graph below (Figure 7) illustrates the parallel trends for the treated (red line) and control (blue line) groups across the two periods: 2019 (pre-treatment) and 2023 (post-treatment).

In the pre-treatment (2019), the treated and control groups follow almost parallel trends, supporting the parallel trends assumption. This indicates that the groups had similar growth trajectories before the treatment (AI adoption).

In the post-treatment (2023), a divergence in the trends is visible, with the treated group showing a slightly higher increase than the control group. This suggests a potential treatment effect caused by AI adoption.

Table 1. The results for fixed effects difference in difference model estimation using median, mean and third quartile criteria.

Variable	Metric
	Median	Mean	Q3
interaction_term	0.3564695 ** [0.1464747]	0.374747 ** [0.1531026]	0.3558718 ** [0.1731414]
log_firm_size	0.7178091 *** [0.2048316]	0.6643637 *** [0.1893067]	0.6390377 *** [0.1894297]
log_own_capitals	0.3736364 ** [0.1766719]	0.4484532 *** [0.141167]	0.4682482 *** [0.1389942]
Cons	8.99362 *** [3.204512]	7.811036 *** [2.529322]	7.56702 *** [2.48518]
sigma_u	1.6946254	1.7208399	1.702649
sigma_e	0.3417927	0.3482887	0.3550451
Rho	0.9609104	0.9606483	0.9583292
R-squared (within)	0.7247	0.7141	0.7029
R-squared (between)	0.3527	0.3504	0.3478
R-squared (overall)	0.3620	0.3581	0.3554
F	26.88 ***	26.46 ***	25.48
Number of obs	54
Number of groups	28

Source: Own processing. Note: **, *** represent 10%, 5% and 1% significance level. [] represent the robust standard errors of the coefficients. The variables used within the model are as it follows: interaction_term: This is the Difference-in-Differences (DiD) estimator. It represents the coefficient β3 from the FE DiD model and captures the average treatment effect of AI adoption on the dependent variable; log_firm_size is the natural logarithm of the number of employees; used as a control variable to account for differences in firm size; log_own_capitals is the natural logarithm of the firm’s own capital (equity); included to control for financial strength.

presents consistent and statistically significant treatment effects for each threshold. The interaction term (β₃) is between 0.3559 and 0.3747, which proves that AI adoption results in a significant turnover increase.

The results provide several noteworthy findings on the relationship between AI adoption and firm productivity. First, company size proves to be a significant determinant of economic output. In particular, a 1% change in the firm size (in employees) is correlated with a 0.64% to 0.72% increase in turnover, which shows that bigger firms have a better chance of extracting the benefits of AI integration.

Second, financial capacity is a major determinant. As reported, a 1% increase in equity leads to a rise in turnover of 37.4–46.8%, reaffirming the fact that equity plays a vital role in supporting the introduction and adoption of AI by corporations through stronger capital structures.

The model is good at explaining the variability in turnover, with within R² values of over 0.70, indicating that the included variables capture an important portion of the variation in turnover. Furthermore, the rho coefficients (ρ > 0.95) indicate that the majority of the variance is firm-specific, which justifies the inclusion of fixed effects to control for unobserved heterogeneity between companies.

Among the classification metrics tested, the mean threshold yields the maximum estimated treatment effect (β₃ = 0.3747), implying that it assigns AI adoption as the largest contributor to turnover growth. Nonetheless, the median was chosen for the main analysis, as it balances a more even distribution of firms and moderates the risk of overestimating the treatment effect while maintaining robustness. Using the coefficient β₃ from the preferred specification and applying the transformation:

$$\mathrm{A}{\mathrm{I}}_{\mathrm{s}\mathrm{h}\mathrm{a}\mathrm{r}\mathrm{e}}=\mathrm{e}\mathrm{x}\mathrm{p} \mathrm{e}\mathrm{x}\mathrm{p}\left({\beta }_3\right)-1$$

We estimated that 42.8% of turnover among treated firms in 2023 can be attributed to AI-related activities. In absolute terms, this translates into about 19.94 billion lei, providing a more concrete measure of AI’s benefits to firm-level economic performance in energy in Romania. Next, we apply this value within the context of the Leontief Input–Output theory to assess the indirect impact of AI adoption on the wider economy and society, such as its contribution to GDP growth and employment generation throughout the entire Romanian economy.

4.5. The Estimated Economic and Employment Effects of AI Adoption in the Energy Sector: A Leontief Input–Output Approach

Building on the firm-level analysis of AI adoption and its estimated contribution to company performance, this section scales the results to the macroeconomic level using the Leontief Input–Output Model. By incorporating the AI-attributable share of turnover identified in the previous Difference-in-Differences analysis, we assess how AI adoption in the energy sector affects GDP and employment across the broader economy.

The economic effects of AI adoption in the energy sector were assessed using the Leontief Input–Output Model, which quantifies how changes in demand within one industry propagate across the entire economy. This model accounts for three types of effects:

• direct effects: increased output in the AI-adopting energy sector itself,
• indirect effects: increased activity in upstream and downstream sectors,
• induced effects: demand stimulated by household income gains resulting from the first two effects.

In the case of the production and supply of electricity, heat, gas, hot water, and air conditioning sector, the Type II GDP multiplier is 2.84. This means that for every 1 RON of final demand, the total GDP impact is 2.84 RON, comprising 0.39 RON from direct effects, 1.96 RON from indirect effects, and 0.57 RON from induced effects.

Applying this multiplier to the AI-related demand increase in 2023, we estimate that AI adoption in this sector contributed 3.54% to Romania’s GDP, broken down as (

Table 2. The Impact of AI Adoption in the Energy Sector on GDP in 2023 (%).

	Production and Supply of Electricity, Heat, Gas, Hot Water, and Air Conditioning
Direct contribution	1.24
Indirect Contribution	1.57
Induced Contribution	0.72
Total Contribution	3.54

Source: Own processing.

):

Therefore, the direct contribution of the energy sector to GDP formation in 2023 was estimated at 1.24%. The indirect impact arises from the interaction of sectors involved in energy production and supply activities with upstream and downstream industries. Therefore, the indirect contribution of the energy sector to GDP was 1.57%. When considering monetary flows in and out of households and their effects on economic sectors, the induced contribution of the energy sector to GDP in 2023 is estimated to be 0.72%. The total estimated impact of the energy sector on GDP in 2023 is 3.54%.

These results highlight the strong ripple effects of AI integration, not only improving the efficiency of the sector itself but also stimulating broader economic activity through supply chain and consumption channels.

The employment impact of AI adoption also reflects a mix of contraction and expansion across different areas of the economy. According to the Leontief employment multiplier, a 1000 RON increase in final demand leads to the creation of 4.90 new jobs, broken down as: 0.86 jobs directly in the energy sector, 3.41 jobs indirectly in related industries, 1.49 jobs induced by increased consumption in the broader economy.

Concerning employment, the incorporation of AI is expected to create new opportunities, particularly in advanced roles such as AI specialists, data analysts, and engineers who design and implement these technologies. Jobs are created indirectly in upstream sectors, such as the production of renewable energy technology and the creation of AI systems, as well as in downstream sectors that gain from improved and more efficient energy systems. Induced effects arise as higher incomes in these industries lead to job growth in consumer-facing sectors like retail and hospitality.

The results presented in

Table 3. The Impact of AI Adoption in the Energy Sector on Employment in 2023.

	Production and Supply of Electricity, Heat, Gas, Hot Water, and Air Conditioning Thousands of Persons
Direct contribution	−22.35
Indirect Contribution	53.88
Induced Contribution	33.49
Total Contribution	65.03

Source: Own processing.

quantify the employment effects of AI adoption in the energy sector in 2023, capturing its direct, indirect, and induced contributions.

The direct contribution of AI adoption in the energy sector led to a net loss of 22.35 thousand jobs, primarily due to automation and efficiency improvements that reduced the need for traditional roles. Although new high-skilled positions—such as AI specialists, engineers, and technicians—have emerged, they have not fully offset the displacement of conventional jobs.

In contrast, the indirect contribution amounted to +53.88 thousand jobs generated in upstream industries such as manufacturing, technology, and consulting that support AI deployment. Additionally, +33.49 thousand jobs were created through induced effects, as increased household incomes stimulated demand in consumer-facing sectors like retail and services.

Overall, the net employment impact was positive, with a total gain of +65.03 thousand jobs. These results underscore the complex dynamics of AI integration: while it displaces labour within core sectors, it simultaneously drives employment growth in adjacent and downstream industries.

5. Discussions

This paper provides novel empirical insights into the economic and social implications of AI adoption in the Romanian energy industry. Utilising NLP-based AI Adoption Scores from media coverage, we perform firm-level econometric analysis to provide a holistic framework for pinpointing legitimate AI adopters while estimating the economic value of their adoption paths. The raw visibility metric and the two credibility-adjusted versions penalised by the MBSs together create a dual-scoring approach that allows for a more nuanced classification of firms—one that captures not only who is engaging with the technology in public discourse but also the quality of the narratives behind those claims.

The Difference-in-Differences (DiD) estimates verify that AI technology adoption relates positively and significantly to company performance. According to the preferred specification, over 42.8% of turnover in firms that adopt AI can be attributed to AI-related activities. These are consistent with international literature [47,48], which highlights the importance of digital technologies for increasing productivity, flexibility and competitive advantage.

At the level of the economy as a whole, the application of Leontief multipliers indicates that AI adoption in the energy sector contributed 3.54% to Romania’s median GDP in 2023 and resulted in a net employment gain of over 65,000 jobs despite direct labour displacement within the sector. These findings support the notion that AI is a systemic catalyst—not just improving firm-level productivity but also driving economic activity at every level: upstream, downstream, and at the points of consumption.

From a methodological perspective, the stability of firm classification across all score versions (raw and penalised) strengthens the reliability of our identification strategy. It ensures that the treatment effects estimated in the FE DiD analysis are not artefacts of media bias.

Our results should also be interpreted with certain caveats. While the FE DiD framework provides robust evidence of the performance impact of AI adoption, we could not fully explore heterogeneity across firm size or subsectors due to the limited sample. Larger companies appear to drive the aggregate effects, but future extensions with more granular data could reveal whether SMEs or specific energy subsectors (e.g., renewables versus traditional utilities) experience different adoption trajectories. On the macro side, input–output multipliers were applied based on the 2021 symmetric table, which is the most recent official dataset.

Second, our approach relies on media-derived signals, which are subject to narrative bias. To address this, we developed the Misinformation Bias Score adjustment, explicitly penalizing speculative or exaggerated reporting. We also attempted to complement these media signals with alternative data sources, including LinkedIn job postings for AI-related positions, patent databases, and corporate sustainability reports. However, in the Romanian energy sector such information was either absent or too rudimentary to be systematically integrated. This highlights the need for richer and more diversified datasets in future research, which would allow triangulation and validation of the media-based adoption scores.

Finally, our policy recommendations should be viewed not only in terms of enabling adoption but also in safeguarding its inclusiveness and credibility. In addition to practical measures such as targeted reskilling programs, fiscal incentives for SMEs, and investments in regional digital infrastructure, there is a pressing need to integrate ethical AI governance principles. This includes transparent reporting standards, the introduction of mandatory AI disclosure audits, and alignment with EU-level frameworks to curb risks of greenwashing and ensure accountability. Such measures would help prevent uneven adoption from deepening regional or social inequalities, while reinforcing trust in the sector’s digital transformation.

These results are relevant to public policy. They argue that driving the adoption of AI requires more than public awareness campaigns or isolated investments. These benefits can be scaled up and made inclusive of targeted interventions—tax incentives, digital infrastructure development, and workforce reskilling, among others. Moreover, our dual-lens scoring framework can assist policymakers in distinguishing firms that are merely AI “signal boosters” versus those that, through credible implementation, are advancing transformative technologies.

However, some limitations should be mentioned. The firm-level sample, while representative in terms of turnover (87% of sectoral activity), is relatively small (36 companies). The media-based AI Adoption Score may not always reflect implementation; future research should consider triangulating this score with survey or administrative data. Similarly, input–output multipliers assume national coefficients, which may not fully reflect the dynamic shifts introduced by AI adoption at the sectoral level.

Despite these constraints, the methodology proposed offers a replicable and scalable framework for analysing AI adoption in strategic sectors. As Romania and other emerging economies navigate the twin transitions of digitalisation and decarbonization, understanding how technologies, such as AI, reshape productivity, labour, and value creation will be essential for designing inclusive, forward-looking industrial strategies. In short, our findings show that AI adoption—when buttressed by credible communication and effective deployment—can drive strong returns for both firms and the economy. However, fully realising its potential will require coordinated policy action to address adoption gaps, infrastructure weaknesses, and workforce transitions. These results have clear implications for public policy that will enable digital infrastructure, incentivise adoption, and mitigate labour displacement—issues addressed further in the Conclusions.

6. Conclusions and Policy Implications

We present a new multi-level analysis of AI adoption in Romania’s energy system that combines natural language processing with econometric modelling and input–output method to estimate both microeconomic outcomes and wider economic and social effects. The results improve our understanding of how AI adoption develops across firms and scales up to affect national economic metrics.

Hence, at the firm level, the adoption of AI technologies correlates with significant performance benefits. According to the Difference-in-Differences model, AI-related activities are responsible for up to 42.8% of turnover for the companies that have adopted AI. For all classification thresholds, these effects are robust to the control of media bias using Misinformation Bias Scores (MBSs), further validating the AI Adoption Score as a valid proxy for digital engagement. In absolute terms, this translates to approximately 19.94 billion lei in turnover attributable to AI among Romanian energy firms in 2023.

The mean threshold yields the highest treatment effect estimate (β₃ = 0.3747), suggesting that firms classified as AI adopters experienced an average turnover increase of approximately 37.5%. Firm size and equity were also statistically significant predictors of performance, with a 1% increase in firm size associated with a 0.64–0.72% increase in turnover, and a 1% increase in own capital linked to a 37.4–46.8% increase in performance.

At the macro level, when extrapolated through the Leontief Input–Output model, the adoption of AI technologies in the energy sector represents an estimated 3.54% of the national GDP in 2023. Importantly, while direct employment within the sector fell as a result of automation (−22.35 thousand jobs), the economy saw a net increase of more than 65,000 jobs, primarily through indirect and induced effects across adjacent and consumer-oriented industries.

Our findings suggest that we should view AI adoption as another strategic lever for economic transformation rather than thinking about digitization as simply a technological upgrade inside firms. But the benefits aren’t evenly distributed, and there are structural barriers (skill shortages, implementation costs and gaps in digital infrastructure) that may restrict wider adoption. This calls for targeted public policy responses.

Among the main policy recommendations for enabling AI Adoption in the energy sector, we can mention:

• Provide tax credits, grants, or innovation vouchers to support AI investments, especially in areas such as predictive maintenance, grid optimization, and demand forecasting.
• Fund technical education and lifelong learning programs focused on AI, data analytics, and digital technologies relevant to the energy sector.
• Establish AI pilot zones and testbeds—for example, for smart grids or energy trading platforms—through collaboration among energy providers, tech firms, and research institutions.
• Incentivize disclosure of AI strategies in corporate ESG and digitalization reports to improve policy targeting and promote best practices.
• Support the development of AI hubs in less digitally developed regions through infrastructure investments, startup support, and local partnerships.

In conclusion, this study highlights the transformative—albeit unbalanced—potential of AI adoption across critical sectors and suggests that more concerted and evidence-based policies are needed to ensure that digital innovation serves not only as a driver of economic growth, but a tool for inclusive and sustainable development.