Building Trust in Large Language Models: Navigating the EU AI Act, Global Standards, and Sustainability Challenges

Introduction

Like an invisible glue holding the world together, our lives represent an ecosystem built on trust. Trust is the bedrock of all personal and professional relationships, but it is also crucial in a citizen-to-government relationship, where governments are trusted to enforce laws fairly and to protect rights. Trust sustains the legitimacy of institutions and the social contract, as well as scientific advancements, business innovations and large-scale projects. People adopt new technologies, products, and practices because they trust the creators or systems behind them.

Large Language Models (LLMs), a type of artificial intelligence (AI) that understands and generates human-like language by processing vast amounts of data, like GPT-4, LaMDA, and LLaMA, make no exception. Building trust in LLMs is crucial not only for their widespread adoption but also for ensuring they are used responsibly and effectively. Additionally, building trust in LLMs is becoming increasingly important the more we realise their immense potential to transform entire sectors – from assisting doctors in diagnosing diseases by summarising complex medical literature to powering customer service chatbots that handle millions of interactions daily. Without trust, the benefits of LLMs are overshadowed by risks — misleading outputs, encoded biases, and environmental costs of training massive models.

The lack of a clear understanding of their internal workings further erodes trust, leading stakeholders, from end-users to policymakers, to question their reliability and integrity. Can we trust an AI-powered legal assistant to provide accurate, unbiased legal advice? Should a journalist rely on AI-generated content when misinformation spreads so easily? Can policymakers confidently use LLMs to analyse economic trends without worrying about hallucinated data?

The consequences of misplaced trust in LLMs can be significant. Misinformation generated by AI could manipulate public opinion during elections. A biased algorithm in hiring tools could reinforce discrimination in job recruitment. The high energy consumption of training and running LLMs raises sustainability concerns, leaving a growing carbon footprint that contradicts global climate goals. Additionally, the “black box” nature of these models makes it difficult for users to understand how decisions are made, leading to scepticism and resistance.

As LLMs become more integrated into critical domains, such as law, healthcare, education, and finance, ensuring their reliability, transparency, and ethical use becomes ever more necessary. This paper explores some of the efforts to build trust in LLMs under the European Union Artificial Intelligence Act (EU AI Act) and other regulations, frameworks, guidelines or projects related to building trust in LLMs. In the end, we will discuss the next steps necessary to establish trust in LLMs, ensuring that their benefits outweigh their risks and that they contribute positively to society.

The Need to Build Trust

In 2023, the OECD launched its AI Incident Monitor (AIM) to follow artificial intelligence (AI) incidents reported in reputable international media. AIM uses machine learning models to identify and classify incidents. Between 2022 and 2023, the number of AI incidents reported increased by approximately 1278%, coinciding with the mainstreaming of generative AI.

AI incidents increased dramatically with generative AI

Meanwhile, the Open Worldwide Application Security Project (OWASP), a nonprofit foundation that works to improve the security of software, created the OWASP Top 10 for LLM Applications list. This, significant undertaking, built on the collective expertise of an international team of nearly 500 experts with over 125 active contributors, further reflects the urgent need to establish security best practices as AI adoption accelerates.

With numerous risks surrounding LLMs becoming more apparent, building trust in these models has proven to be a complex challenge. A few key concerns further drive this complexity. Firstly, LLMs operate as black boxes, even to their developers. Their architecture and training processes, often proprietary and complex, are not easily interpretable. Furthermore, LLMs exhibit “capability unpredictability,” where their strengths and weaknesses are not fully understood until tested on specific inputs. This unpredictability is compounded by emergent behaviours that arise as models scale up, making it difficult for developers to anticipate all potential uses or misuses.

Secondly, LLMs are known to encode and perpetuate biases present in their training data. This undermines their utility in high-stakes applications, but beyond technical flaws, LLMs have societal implications. They may generate harmful content, amplify stereotypes, or disrupt labour markets by automating tasks previously done by humans.

Overall, we can say that transparency and predictability, cornerstones of trust, are difficult to achieve with LLMs, and the “move fast and break things” approach of the rapid deployment of LLMs in a variety of sectors worldwide has exacerbated mistrust, as stakeholders struggle to keep up with the implications of new models and applications.

To address these challenges, trust in LLMs must be approached holistically, integrating technical transparency with ethical considerations and societal impact.

Transparency. Providing clear, tailored explanations of how LLMs work, their limitations and the reasoning behind their outputs is essential. However, these efforts must be meaningful and actionable for stakeholders.

Legal and Ethical Frameworks. Incorporating principles like fairness, accountability, and non-discrimination, as outlined in frameworks like Assessment List for Trustworthy Artificial Intelligence (ALTAI), ensures that LLMs align with societal values and regulations, such as the EU AI Act, can be a framework guiding LLMs to build their trustworthiness at EU level.

Sustainability. Addressing economic, social, and environmental impacts can create a foundation for long-term trust. For example, improving energy efficiency, reducing bias, and designing models for equitable outcomes can mitigate concerns.

By focusing on these aspects, we can move beyond technical achievements to ensure that LLMs serve humanity in trustworthy and sustainable ways.

Legal and Ethical AI Landscape

Efforts to build trust in LLMs begin with robust ethical guidelines and regulatory oversight. The ALTAI framework provides a comprehensive checklist for ensuring ethical AI deployment. It emphasises seven critical principles.

1. Human agency and oversight: This requires that AI systems should both: 1) act as enablers for a democratic, flourishing and equitable society by supporting the user’s agency; and 2) uphold fundamental rights, which should be underpinned by human oversight.
2. Technical robustness and safety: This requires that AI systems are developed with a preventative approach to risks and that they behave reliably and as intended while minimising unintentional and unexpected harm as well as preventing it where possible.
3. Privacy and data governance: Prevention of harm to privacy also necessitates adequate data governance that covers the quality and integrity of the data used, its relevance in light of the domain in which the AI systems will be deployed, its access protocols and the capability to process data in a manner that protects privacy.
4. Transparency: This encompasses three elements: 1) traceability, 2) explainability and 3) open communication about the limitations of the AI system.
5. Diversity, non-discrimination, and fairness: AI systems should be user-centric and designed in a way that allows all people to use AI products or services, regardless of their age, gender, abilities or characteristics.
6. Societal and environmental well-being: Sustainability and ecological responsibility of AI systems should be encouraged, and research should be fostered into AI solutions addressing areas of global concern.
7. Accountability: When unjust or adverse impacts occur, accessible mechanisms for accountability should be in place that ensure an adequate possibility of redress.

Even though ALTAI is not a solely focused framework for LLMs, these principles guide developers and organisations in creating systems that respect human values while delivering reliable and equitable outcomes.

On the other hand, regulatory frameworks like the EU AI Act complement these ethical considerations by imposing structured governance. The EU AI Act represents one of the most ambitious regulatory frameworks for governing AI technologies. At its core, the Act categorises AI systems based on their potential risks to individuals, society, and fundamental rights. LLMs, due to their general-purpose nature and significant societal impact, are addressed within two critical categories.

1. High-Risk AI Systems. The Act designates systems as high-risk if they pose a significant threat to health, safety, or fundamental rights. Examples include AI used in critical infrastructure, law enforcement, and employment processes. High-risk systems are subject to stringent requirements such as extensive risk management, transparency and accountability, robustness and security and human oversight.
2. General Purpose AI (GPAI). Recognising the unique challenges posed by systems like LLMs, the Act introduces provisions for GPAI that address the broad applicability of such systems, emphasising the need for adaptability across various contexts, traceability to promote transparency and harmonised standards across industries and applications.

EU AI act risk-based approach, via the European Commission

The EU AI Act is also a regulatory framework addressing the whole AI system, but it provides specific obligations for LLMs, helping stakeholders to build trust in LLMs. Here are some obligations and requirements related to LLMs.

• Article 53: Obligations for Providers of GPAI Models: Maintain and update the model’s technical documentation, including training, testing, and evaluation results, provide updated information and documentation to AI system providers integrating the GPAI model, Implement a policy ensuring compliance with EU copyright and related rights laws, publish a detailed summary of the training data content and cooperate with the European Commission and national authorities as needed.
• Article 55: Obligations for Providers of GPAI Models with Systemic Risk: Conduct model evaluations using standardised protocols and state-of-the-art tools, including adversarial testing, to identify and mitigate systemic risks, assess and address potential systemic risks at the EU level arising from the development, market placement, or use of the model, monitor, document, and report serious incidents and corrective measures to the AI Office and national competent authorities without undue delay and ensure adequate cybersecurity protection for the model and its physical infrastructure.
• High-Risk AI System
  • Article 10 (Data Governance and Bias Mitigation): High-risk AI systems must be developed using high-quality datasets, with data management covering design, collection, preparation, evaluation, and bias prevention. Datasets must be representative, accurate, and reflect the specific context of use. Sensitive data may be processed for bias correction under strict conditions, including minimal use, security measures, restricted access, and deletion. If no AI model training is involved, these requirements apply only to testing datasets.
  • Article 13 (Transparency and Explainability): High-risk AI systems must ensure transparency and include instructions covering provider details, system capabilities, risks, oversight measures, maintenance, and log management, enabling deployers to interpret and use outputs safely and effectively.
  • Article 43 (Conformity Assessment): High-risk AI systems require conformity assessments based on harmonised standards or specifications. Depending on the context, providers use internal control or involve notified bodies. Substantial modifications need reassessment unless predefined. The Commission can update procedures to address risks or technical progress.

Beyond frameworks and regulations at the EU level, there are many ongoing efforts to build trust in LLMs. At the G7 Hiroshima Process on Artificial Intelligence in 2023, the OECD published Towards a G7 Common Understanding on Generative AI report which provides an overview of national and regional initiatives to address risks related to generative AI, synthesised in Table 1 below.

Table 1.  Overview of National and Regional Initiatives to Address Risks Related to Generative AI

However, building trust extends beyond ethics and regulation. In a world increasingly concerned with sustainability, LLMs must also address their economic, social, and environmental impacts.

Addressing the Elephant in the Room. Societal, Economic and Environmental Impacts

While ethical and regulatory frameworks are foundational for trust, they are not sufficient on their own. The growing focus on sustainability—spanning economic, social, and environmental dimensions—adds a critical layer of responsibility for LLMs.

• Economic Sustainability. The efficiency and cost-effectiveness of LLMs are crucial for their broader adoption. This includes addressing the high computational resources required for training and deploying these models, ensuring they deliver value without excessive economic burdens.
• Social Sustainability. Trust in LLMs also hinges on addressing risks of bias and discrimination. If unchecked, LLMs may perpetuate or even amplify existing societal inequities. Solutions such as diverse datasets and fairness testing are essential to ensure equitable outcomes for all users.
• Environmental Sustainability. Finally, the environmental impact of LLMs, including their significant carbon footprint, must be a priority. Advancements in green AI practices and energy-efficient model designs can help mitigate these effects.

Three Pillars of United Nations’ Sustainable Development Goals (SDGs)

Global organisations like the UN and UNESCO, as well as local governments, companies, and institutions, play vital roles in shaping guidelines and policies that address these sustainability challenges and in offering a holistic approach to ensuring that LLMs not only align with present-day ethical standards and legal requirements, but also contribute positively to a sustainable and inclusive future.

UNESCO, for example, has been a leading advocate for the ethical use of AI, emphasising transparency, accountability, and inclusivity. Its Recommendation on the Ethics of Artificial Intelligence outlines principles that resonate with both the SDGs and regulatory frameworks.

• Human-centred Design. Ensuring AI serves humanity by promoting education, cultural diversity, and equality aligns with SDGs like Quality Education (SDG 4) and Reduced Inequalities (SDG 10).
• Environmental Responsibility. UNESCO’s emphasis on minimising AI’s environmental impact supports SDG 13 (Climate Action) and SDG 15 (Life on Land).
• Global Collaboration. Encouraging international cooperation on AI ethics aligns with SDG 17 (Partnerships for the Goals), fostering a shared commitment to sustainability.

Making such connections is crucial for several reasons. Firstly, the SDGs establish shared global priorities, making them a powerful lens for evaluating the societal and environmental impact of LLMs. By addressing issues that resonate across nations and cultures, aligning LLMs with responsible innovation guidelines and frameworks demonstrates a commitment to universal values, enhancing their acceptance and legitimacy.

Secondly, trust in LLMs cannot be built solely on technical excellence or regulatory compliance. A holistic perspective targeting sustainability is essential for fostering trust among diverse stakeholders, and the SDGs emphasise this interconnectedness, ensuring that LLMs are not only efficient but also socially equitable and environmentally responsible.

Economic Sustainability

A key factor in enhancing trust in LLMs is enhancing its economic sustainability beyond immediate profit. And not surprisingly so – the long-term viability and trustworthiness of LLMs are crucial for investments, but socially speaking, the economic footprint of LLMs, encompassing efficiency, cost-effectiveness, and accessibility, becomes ever more crucial for trust. Fostering inclusivity and innovation beyond slogans, and aligning LLMs with global economic goals, is something we estimate will become a linchpin in economic sustainability.

The United Nations’ Sustainable Development Goals (SDGs) provide some actionable recommendations in this regard.

Efficiency and Cost-Effectiveness of LLMs

Optimising the efficiency of LLMs directly supports SDG 8. Decent Work and Economic Growth and SDG 9. Industry, Innovation, and Infrastructure. By reducing computational costs, businesses can deploy LLMs more broadly, democratising access to advanced AI tools and driving economic growth across diverse sectors. This fosters innovation while ensuring inclusivity in global markets. Furthermore, cost-effective deployment enables a wider range of stakeholders—from small enterprises to public institutions—to leverage the transformative potential of LLMs without incurring prohibitive expenses.

Challenges of Scalability and Resource Allocation

The scalability of LLMs, especially for small- and medium-sized enterprises (SMEs), remains a significant challenge, as found by S&P’s market analysis commissioned by WEKA. To expand further, challenges often faced by smaller businesses when considering LLMs span across several domains.

Firstly, smaller businesses often face significant challenges in accessing the computational resources required for training and deploying LLMs. One of these factors is the infrastructure or cloud service costs. Hardware is still expensive to acquire and maintain, while renting cloud resources at the scale required for LLMs, could be financially prohibitive. Secondly, let us not forget also the lack of in-house expertise to optimise the use of computational resources effectively, leading to inefficient utilisation of available resources and further increasing costs, which has been identified by McKinsey as one of the key limitations for small businesses in AI adoption. Addressing this issue aligns with SDG 10. Reduced Inequalities, as it enables equitable access to AI technologies, allowing smaller players to compete in an AI-driven economy.

Innovations for Optimising LLM Efficiency

On the other hand, technological advancements can support us in ensuring the responsible use of computational resources while maintaining high performance, fostering economic trust and stability. For instance, parameter-efficient tuning, adaptive compression, and low-cost inferencing methods promote SDG 12. Responsible Consumption and Production.

Techniques like model pruning and quantization can significantly reduce the energy and hardware demands of LLMs, making them more accessible to a broader range of users. Similarly, scalable frameworks for distributed computing enable efficient resource utilisation, further enhancing the cost-effectiveness of LLMs.

Social Sustainability

We can hardly speak about progress when discussing technological advancements unless said technologies work for the betterment of humanity. Human-centred technological advancements work within our society, by putting humans at the heart of technological progress and ensuring justice, equity, and inclusivity across diverse populations.

We live in a world already ridden by bias and inequality. The potential for LLMs to exacerbate social inequities hinders trust and stops adoption. Focusing on deliberate design, rigorous oversight and a commitment to ethical development, we can support LLMs’ alignment with global development goals and establish LLMs as powerful tools for positive social change and inclusive future.

Addressing Discrimination and Bias in LLMs

Bias in LLMs often originates from the training data, which may reflect societal stereotypes or historical injustices. If unchecked, biased outputs can harm marginalised groups by reinforcing discriminatory practices or limiting opportunities. For example, biased hiring tools or content moderation systems can disadvantage individuals based on gender, ethnicity, or socioeconomic status.

Mitigating bias in LLMs supports SDG 5. Gender Equality and SDG 10. Reduced Inequalities by ensuring that outputs are fair, equitable, and representative of diverse populations. Ensuring representative datasets, implementing rigorous testing protocols to identify and mitigate bias at every stage of development and collaborating with diverse communities are a few steps that can support us adopt a global mindset and incorporate diverse concerns and perspectives into AI systems.

Highlighting Risks of Perpetuating Social Inequalities

Without intervention, LLMs risk amplifying existing inequalities, such as by limiting opportunities for underrepresented groups or reinforcing systemic biases. These risks are particularly concerning in high-stakes applications such as education, healthcare, and law enforcement. For example, predictive policing tools powered by LLMs have faced criticism for disproportionately targeting marginalised communities (for an analysis of the integration of AI in the judicial decision-making process, see here).

Addressing these issues ensures alignment with SDG 1. No Poverty and SDG 16. Peace, Justice, and Strong Institutions to foster a fairer, more inclusive society.

Frameworks and Methodologies to Mitigate Bias

Techniques such as diverse dataset curation, fairness testing, and ethical auditing align with SDG 4. Quality Education and SDG 17. Partnerships for the Goals. By engaging interdisciplinary teams and leveraging diverse perspectives, the development of bias-mitigation frameworks can set a global standard for equitable AI.

A few actions to support this goal are conducting regular audits to ensure compliance with ethical standards and address any emerging biases, using quantitative and qualitative methods to evaluate the fairness of model outputs across different groups and of course offering enhanced transparency in model development.

Environmental Sustainability

Training and deploying of LLMs is not without its environmental impact. Training state-of-the-art LLMs requires vast computational resources, leading to significant energy consumption and greenhouse gas emissions. For instance, training a single large-scale LLM can generate carbon emissions comparable to the lifetime emissions of several cars. This environmental cost is exacerbated by the increasing scale and complexity of models, as well as their widespread deployment across industries.

Moreover, the environmental impact extends beyond carbon emissions to include e-waste generated by frequent hardware upgrades and the strain on natural resources used in the production of computational infrastructure. Addressing these challenges is critical to ensuring that the benefits of LLMs are not offset by their environmental harm

The carbon footprint of training and deploying LLMs is significant, conflicting with SDG 13. Climate Action. Energy-intensive processes contribute to greenhouse gas emissions, underscoring the need for sustainable AI practices to mitigate environmental harm.Such practices, commonly implementing energy-efficient models, green computing practices, and renewable energy sources support SDG 7. Affordable and Clean Energy and SDG 15. Life on Land. 

Key strategies to tackle the environmental impact of AI include:

1. Energy-Efficient Model Design. Through techniques such as parameter-efficient tuning and sparse model architectures, the computational demands of training and inference are reduced. Additionally, streamlined algorithms for inferencing minimise energy usage without compromising performance.
2. Green Computing Practices. Leveraging distributed computing across multiple, optimised nodes reduces energy consumption by improving computational efficiency. This is a common green computing practice, and it includes dynamically allocating computational resources based on real-time demand, which therefore minimises unnecessary energy use.
3. Renewable Energy Integration. Transitioning data centres to renewable energy sources, such as solar or wind power, significantly reduces the carbon footprint of AI operations, while utilising region-specific renewable energy options enhances sustainability and aligns with local environmental goals (for a discussion of data centres in Ireland, see here).
4. Optimised Model Architectures:
   1. Pruning and Quantization: Removing redundant parameters and compressing model weights reduced computational intensity.
   1. Neural Architecture Search (NAS): Employing automated tools to design energy-efficient architectures ensures optimal resource utilisation.
5. Lifecycle Management, such as circular economy practices and sustainable manufacturing, encourages recycling and repurposing obsolete hardware to reduce e-waste and the demand for new resources, and encourages manufacturers to adopt green practices in producing computational infrastructure supporting a holistic approach to sustainability.

Last but not least, an all-encompassing strategy for environmental sustainability, as well as social and economic sustainability alike, is a multisectoral and multidisciplinary collaboration. This also includes advocacy for establishing standards for energy-efficient AI development to promote consistency and accountability as well as lobbying for subsidies or tax benefits for companies that adopt sustainable AI practices in order to encourage wider implementation.

Building Bridges Between Goals, Regulation, and Ethics

Trust in LLMs is not merely a regulatory milestone; it is a deeply human construct shaped by fears, preconceived notions, misinformation, and cognitive frameworks. While regulatory frameworks like the EU AI Act provide essential guardrails, they address only part of the equation. Building trust requires addressing these intimate human processes and creating a holistic approach that goes beyond compliance to foster genuine understanding and confidence.

At the heart of this holistic approach lies the recognition that trust must be earned through multiple interconnected efforts.

There is a need to embed sustainability in design: By integrating economic, social, and environmental considerations at every stage of LLM development, organisations can ensure that their systems align with global priorities such as the UN’s SDGs. This proactive incorporation of sustainability fosters trust by demonstrating responsibility and foresight. For example:

• Economic Sustainability: OpenAI has explored parameter-efficient models that reduce the cost of training, enabling broader accessibility for SMEs.
• Social Sustainability: Google’s AI principles emphasise fairness and inclusivity, exemplified by efforts to mitigate bias in models like LaMDA through diverse dataset curation.
• Environmental Sustainability: Microsoft’s commitment to carbon-negative AI operations includes leveraging renewable energy-powered data centres for training large-scale models such as OpenAI’s GPT series.

There is a need to leverage multi stakeholder collaboration: Governments, international organisations, academia, and private entities must collaborate to address complex challenges and establish best practices. Such partnerships ensure that diverse perspectives inform the development of LLMs, making them more inclusive and responsive to societal needs.

• Government and Industry Partnerships: The EU’s Horizon 2020 program funds collaborative AI projects that engage both private companies and academic researchers to develop ethical AI systems.
• Global Forums: The G7 Hiroshima Process on AI fosters international dialogue to address risks associated with generative AI and establish shared guidelines.

There is a need prioritise transparency and accountability: Trust thrives in environments where stakeholders are well-informed and can rely on the integrity of the systems they use. Clear documentation, open communication, and demonstrable accountability reinforce trust by making the development and deployment of LLMs understandable and relatable.

• Documentation Standards: Hugging Face’s Model Cards provide detailed information about datasets, training processes, and limitations for their open-source AI models
• Transparency in Governance: UNESCO’s Recommendation on the Ethics of Artificial Intelligence outlines clear guidelines for ethical AI development and requires member states to report on their compliance.

Beyond these steps, trust-building efforts must actively raise awareness, inform, and educate. Stakeholders—from developers and policymakers to end-users—must be equipped with the knowledge to engage meaningfully with LLMs, dispelling misconceptions and empowering informed interactions. Integrating ethical frameworks into LLM development, even when not mandated by regulation, further underscores a commitment to societal values, bridging the gap between technical innovation and human-centred principles.

Ultimately, trust is not built solely on rules or policies but within the intricate interplay of understanding and belief. By embracing a holistic approach that integrates sustainability, collaboration, transparency, and education, we can create LLMs that are not only powerful tools but also trusted companions in the journey toward a more equitable and sustainable future. This layered and human-centric strategy positions LLMs as catalysts for meaningful progress, earning the confidence of all who engage with them.