What explainable AI is, why it matters and how we can achieve it

AI is reshaping critical sectors of society like healthcare, finance, and justice. From diagnosing diseases, deciding loan approvals, to judicial outcomes, AI’s decisions can deeply affect our lives. But can we trust these systems when their inner workings remain hidden, locked away in complex computational models such as deep neural networks that humans can only perceive as opaque “black boxes”? The need for greater transparency and trustworthiness in AI is becoming increasingly important as these systems become widely deployed, especially in critical sectors.

Explainable AI, in contrast, utilises techniques to clarify how AI models operate, allowing us to trust, verify, and responsibly use these advanced technologies. However, explainable AI is not a perfect solution and faces its own difficulties when attempting to elucidate the machine learning models that underpin AI. The issues surrounding explainability become even more pronounced with large language models (LLMs), which are among the most popular types of AI today. 

For stakeholders such as policymakers, regulators, deployers and end-users of AI technology, it’s helpful to have a basic understanding of explainable AI concepts, methods and approaches, but also its potential and limitations. This can help form realistic demands and raise the explainability standards for AI systems used in future.

Why do we need explainability?

Modern AI can perform impressive tasks, ranging from driving cars and predicting protein folding to designing drugs and writing complex legal texts. Yet, despite these successes, AI systems often operate opaquely, making it challenging to understand or trust their outputs. This lack of transparency isn’t only inconvenient; it poses security, legal, ethical, and practical risks. For instance, an AI system that denies a loan must explain its reasoning to ensure decisions aren’t biased or arbitrary.

Especially when AI is involved in sensitive decisions, such as medical diagnoses and judicial rulings, the necessity for transparency becomes even more pronounced. Without adequate explanations, AI decisions risk violating legal rights, perpetuating biases, or leading to unintended, harmful consequences.

Decisions by AI systems must be understandable, verifiable, and accountable. An increasing number of researchers, regulators, and users recognise that without sufficient explainability, AI cannot become a trusted technology in decision-making processes.

Explainable AI addresses these needs by enabling users and decision-makers to understand the logic behind a model’s operation. It aims to bridge the gap between AI’s performance and our need for understanding its behaviour.

Understanding explainable AI

Explainable AI refers to the ability of an AI model to clearly explain its functioning in a way that humans can understand. This goes beyond technical clarity and involves several related concepts:

• Transparency: Users can access information about the internal workings of the AI system.
• Accountability: AI decisions can be traced, and responsibility can be clearly assigned.
• Traceability: The ability to reconstruct decision-making processes, critical for auditing and oversight purposes.
• Interpretability: Users can easily understand how input data leads to specific outcomes.
• Trustworthiness: AI systems consistently produce reliable, fair, and ethically sound results.

The broader term trustworthy AI is used to describe systems that are predictable, robust, fair, ethically designed, and aligned with social and legal norms. Trust is built on understanding. If users grasp the logic of a system, they are more likely to follow its recommendations. Through explainability, AI evolves from a mysterious technology into a tool that humans can control, oversee, and trust.

It’s also worth mentioning that explainability varies depending on the audience. What a data scientist finds clear may confuse a judge or a regulator. Therefore, explainable AI must translate complex AI operations into understandable explanations tailored for specific audiences, ensuring the practical usability of AI across diverse contexts.

AI’s performance-interpretability trade-off

Many powerful AI models, such as deep neural networks, provide impressive accuracy in tasks like image recognition and natural language processing today. They identify complex patterns from vast data sets without explicit human-programmed rules. However, these advanced capabilities come at a cost: these models do not readily offer explanations for their outputs. They are notoriously difficult to interpret due to non-linear dependencies, large numbers of parameters, and multi-layered structures. As a result, the inner workings are often incomprehensible, even to experts. This combination of high performance and low interpretability thus creates a “performance-interpretability trade-off” dynamic.

This creates a dilemma: should we prefer simpler, less accurate models with transparent operations, or complex, high-performance models that remain opaque? The choice depends significantly on the application’s context. In fields where the consequences are substantial, even marginally lower accuracy may be acceptable if it significantly increases the comprehensibility and trustworthiness of an AI system’s decisions.

Approaches to explainability

Explainability approaches generally fall into two categories:

Prioritising intrinsically interpretable models

Interpretable models, such as linear and logistic regression, decision trees or other simpler rule-based systems, inherently provide transparency because their decision processes are straightforward and easily traceable. They enable users to understand directly how specific data points influence outcomes, facilitating verification and auditing. For explainability purposes, it is therefore advantageous to select these models whenever possible.

Applying post-hoc explanation methods

For more complex models, such as random forest and various neural networks, post-hoc explanations can clarify outputs (e.g., decisions or generated content) after they have been made without interfering with the models themselves. These include:

• Model-specific methods: Tailored explanations for specific model types (e.g., visualisation of decision trees).
• Model-agnostic methods: General techniques applicable across various model types that do not require access to internal structures and rely solely on input/output analysis (e.g. permutation feature importance).
• Local explanations: These focus on individual predictions. Methods such as SHAP or LIME provide clarity on how each input variable contributes to specific decisions.
• Global explanations: Provide insights into overall model behaviour, identifying general rules and decision patterns across the data (again SHAP or LIME); essential for evaluating fairness and regulatory compliance.

Combining these methods often yields the most effective solutions by tailoring explanations to user needs and context.

Explainable AI’s challenges

Despite its potential, explainable AI still faces substantial technological hurdles:

• Inconsistent methods: A lack of standardised evaluation for explainability methods themselves makes it difficult to ensure reliability and reproducibility, making it hard to determine which explanation is “true” or most useful.
• Handling unstructured data: Traditional methods struggle with AI systems that process complex unstructured data like text or images, where context heavily influences meaning, arising not just from individual words but from context, syntax, and implicit references.
• Computational limitations: Running explainability methods on large AI models, which have billions of parameters, such as LLMs, requires significant computational power, making real-time explanations challenging.
• Effective communication: Even accurate explanations become ineffective if they are neither understandable nor actionable for the intended audience (e.g. lawyers, judges, or regulators).

The multidimensional nature of unstructured data is an added challenge, especially for text. This is particularly true for explainability methods and creates a need for innovative approaches to maintain effectiveness.

Explaining LLMs

Today’s widely used LLMs, such as ChatGPT, Claude, or Gemini, are paramount to modern AI applications in generating natural language and present unique challenges. Although highly capable of generating coherent, context-rich content, their size and complexity make it very hard to explain why the model gave a particular answer, connected certain concepts, or excluded others. Determining why they produce certain outputs or biases is very challenging, making it the ultimate “black box”.

LLMs also often reflect biases present in their training data. Without explainability, it is difficult to determine whether generated content is discriminatory, incorrect, or unethical. This means using such models in many contexts is risky, such as judicial, potentially leading to systemic errors or human rights violations.

Explaining text-based outputs is a challenge. Text data are high-dimensional, context-sensitive, and rich in implicit meaning, and conventional explainability methods struggle to capture complex conceptual relationships. For instance, LLMS may produce an accurate legal summary; however, it remains unclear which parts of the original document were deemed relevant and which legal knowledge was implicitly applied.

Since these foundation models offer the underlying technology for the rapidly evolving field of Agentic AI, which will have widespread real-world impact, it is important to assess to what extent current explainability approaches can improve oversight of model behaviour.

To date, there are several approaches, typically categorised by how models are trained and utilised. Zhao et al. (2024) classify explainability techniques for LLMs into two paradigms: 

• Fine-tuning: Involves adapting a pre-trained model on a smaller, domain-specific dataset to enhance task performance (e.g., legal interpretation, medical diagnostics); Here, the explanations focus on how fine-tuning affects attention structures, semantic representation, or key entities and phrases.
• Prompting: Uses specially crafted inputs (instructions, questions, examples) without altering the model itself. Explanatory methods analyse which prompt elements influenced the response or how context shaped the outcome. Since internal representations are not accessible, these explanations often rely on changes to input/output.

Understanding behaviour in either paradigm is difficult due to the highly non-linear architecture of LLMs. Minor changes to input can lead to significant differences in the output, complicating the ability to provide stable, repeatable explanations.

Other experimental methods include attention visualisation, token influence analysis, and contrastive learning to detect key variations in output. Together, these techniques form a rapidly growing research field aimed at revealing LLMs’ inner workings to support reliability, fairness, and regulatory compliance. For more in-depth information, see Zhao et al. (2024).

Explainability as standard practice

As described, explainable AI is a broader paradigm that applies special methods and approaches to improve our understanding of complex AI models. It is a crucial component on the way to trustworthy AI and is therefore linked to concepts such as transparency, accountability, traceability, interpretability and trust in AI. Prioritising explainability sets the foundation for AI technology to become a supportive tool for humanity based on reinforced trust, enhanced fairness, and alignment with legal and ethical norms. 

There are still many explainability challenges for AI, particularly concerning widely used, complex LLMs. For now, deployers and end-users of AI face challenging trade-offs between model performance and interpretability. What is more, AI may never be perfectly transparent, just as human reasoning always has a degree of opacity. But this should not diminish the ongoing quest for oversight and accountability when applying such a powerful and influential technology.

On the contrary, these limitations should motivate a more serious and committed pursuit of explainability. On the path towards efficient, safe and responsible AI deployment, explainability should be a core design principle and become a universal standard that steers future AI research, regulation, and institutional adaptation.