Explainability and Trustworthiness in Large Language Models: Implications for Industry and Policy
Introduction
Large Language Models (LLMs) have achieved remarkable capabilities in natural language generation and understanding, powering applications from virtual assistants to code generation. However, their inner workings remain largely opaque, earning them a “black box” reputation that hinders trust, especially in high-stakes domains [1]. Lack of transparency means users and stakeholders cannot fully understand how or why an LLM produced a given output, which is problematic in contexts like healthcare, finance, or legal decision-making where unexplained errors or biases can have serious consequences [1]. As LLMs make their way into enterprise products and government services, explainability and trustworthiness are no longer optional features but essential requirements. Policymakers are also circling this issue: the EU AI Act and similar regulatory efforts explicitly demand transparency and the ability to audit AI decisions in understandable terms [2].
This article explores explainability and trustworthiness in LLMs through an academic lens, with a focus on industrial and policy implications. We integrate technical depth—discussing architecture-level interpretability, traceable reasoning paths, and uncertainty estimation—with practical considerations for deploying LLMs in enterprise and government settings. We review related work on explainable AI for LLMs, analyze the unique challenges posed by LLM explainability in high-stakes applications, and survey emerging directions such as evidence-bound reasoning, causal inference with graphs, and counterfactual explanations. We also examine regulatory context (e.g., EU AI Act compliance, auditability requirements) to illustrate why explainability is crucial for risk management and accountability. Finally, we propose concrete design strategies for making enterprise-grade LLM systems more explainable, and discuss open research problems and future directions in this rapidly evolving field.
Related Work and Background
Explainable AI (XAI) is a broad research area devoted to making the behavior of AI models interpretable to humans. Traditional XAI techniques include model-agnostic methods like LIME and SHAP for feature attribution, visualization of attention weights, saliency maps, and counterfactual reasoning, among others [3,4]. However, applying these to LLMs is challenging due to the sheer scale (billions of parameters) and the generative, sequential nature of LLM outputs. Standard feature attribution can identify which input tokens influenced a classification outcome, but explaining an open-ended text generation is more complex. Attention-based explanations—highlighting which parts of the prompt the model “attended” to—offer some insight into input focus [4], yet attention alone may not fully trace the model’s reasoning.
Recent research in mechanistic interpretability attempts to reverse-engineer Transformer models by analyzing internal activations and learned representations [5]. This has yielded some insights (e.g., identifying neuron circuits for specific behaviors), but it remains an arduous process and not yet a practical solution for real-world deployments. In summary, while a variety of explainability techniques exist, they were largely developed for smaller or more structured models; LLMs’ complexity and the fluidity of language output push the limits of current XAI methods [6].
A key distinction in related work is between post-hoc explanations and intrinsic interpretability. Post-hoc methods treat the trained LLM as a black box and produce explanations after the fact—for instance, generating a natural language explanation for an LLM’s output or using counterfactual examples to highlight important input changes [7]. Intrinsic interpretability, on the other hand, involves designing models or architectures that are transparent by design. For example, some approaches train LLMs to produce a step-by-step rationale along with their answers (sometimes called self-rationalization), or combine an LLM with an interpretable model like a rules engine or decision tree [8]. These hybrid strategies aim to balance the power of LLMs with the clarity of simpler decision models [8].
Another line of work conditions LLMs to explain other models’ decisions: recent studies show that LLMs can transform complex model outputs into human-readable explanations, essentially acting as an “explainer” for other AI systems [9,10]. This highlights the dual role LLMs can play—not only do they require explanation themselves, they can be tools to generate explanations. The efficacy and faithfulness of such explanations (i.e., whether the explanation truly reflects the model’s reasoning or just sounds convincing) is an ongoing research question.
Contemporary LLMs like GPT-4 and PaLM demonstrate impressive fluency and knowledge, but they also exhibit behavior that undermines trust. One notorious issue is hallucination: LLMs can produce plausible-sounding statements that are entirely fabricated or incorrect. In enterprise applications, hallucinations can be more than embarrassing—they can mislead decision-makers or violate compliance if, say, an LLM-generated report cites non-existent data. Studies have observed that LLMs, when used in complex domains (e.g., IT incident management), often conflate symptoms with causes and produce incorrect explanations for events [11]. Moreover, LLM outputs can be inconsistent; the same question might yield different answers across sessions, or slight changes in phrasing can flip the response. This variability is unacceptable in high-stakes settings that require consistent, repeatable outcomes.
Another concern is overconfidence: LLMs tend to state all outputs in a confident tone, lacking a calibrated notion of uncertainty. A model may assert an answer with unwarranted authority, which is dangerous if users take it at face value. Bias and fairness issues also lurk as trustworthiness concerns—LLMs trained on Internet text have learned societal biases, but without transparency it is difficult to detect or mitigate those biases in specific decisions. In summary, the lack of explainability exacerbates each of these issues: when an error or bias occurs, we have little insight into why it happened or how to prevent it next time. These challenges motivate new research into making LLM reasoning more traceable and reliable.
Challenges of LLM Explainability in High-Stakes Applications
Deploying LLMs in enterprise and government applications brings unique challenges for explainability. In such high-stakes contexts (e.g., medical diagnosis assistance, credit risk assessment, policy drafting, legal analysis), the tolerance for mistakes is low and the demand for accountability is high.
Modern LLMs are extremely large, with thousands of layers and attention heads encoding abstract features. This complexity means traditional interpretable ML techniques often break down—for example, local approximation methods like LIME struggle to faithfully represent a decision boundary in a 175 billion-parameter model [12]. High-stakes decisions typically require detailed justifications, but an LLM’s reasoning is distributed across many internal dimensions that do not have human-understandable meaning. This makes it hard to extract a simple explanation such as “Feature X led to outcome Y.” The scale and depth of LLMs thus pose a fundamental challenge: how to trace a particular decision through the network’s computations.
Unlike rule-based systems or software code, LLMs do not follow explicit IF-THEN logic that one can audit. They perform a complex sequence of vector transformations. While techniques like chain-of-thought prompting encourage the model to generate intermediate reasoning steps in natural language, these steps are themselves model-generated text, not guaranteed to reflect the actual internal decision path. In critical applications, a decision process should be reproducible and reviewable by auditors. Current LLMs offer limited internal traceability—we often can’t pinpoint which training data or which intermediate representation led to a specific output. This is a serious issue for auditability and accountability: for example, if an LLM-based system rejects a loan application or recommends a policing action, regulators may ask for a clear rationale. Without traceable reasoning, we risk “the algorithm said so” as the only answer, which is unacceptable in governed contexts.
Enterprise deployments often face data that differ from the LLM’s training distribution. In high-stakes use, an LLM must handle novel inputs (e.g., a new type of medical report or an evolving financial scenario) reliably. Explainability is challenged under domain shift because when the model errs, it may do so in unpredictable ways. Traditional XAI like feature importance might highlight spurious correlations the model latched onto. For instance, an LLM aiding in legal document review might incorrectly generalize from a pattern in its training data to a new case—without explanation, the end-user won’t know the model’s reasoning was off-track. High-stakes settings amplify the cost of such errors, and lack of explanation makes it difficult for human supervisors to trust the system’s adaptability or to catch mistakes in time.
Explainability is increasingly becoming a compliance requirement (discussed further in the policy section). Regulations and internal governance in sectors like finance or healthcare demand that automated decisions be explainable to affected individuals and auditors. If an LLM cannot provide an explanation, it may simply be disallowed for certain uses. From a liability perspective, if an AI system causes harm (e.g., incorrect advice leading to financial loss or injury), the deploying organization must demonstrate due diligence. Not being able to explain the model’s behavior makes it harder to prove that the system was behaving reasonably or to pinpoint faults. Thus, risk management departments in enterprises are hesitant to green-light LLM solutions without robust transparency and control mechanisms.
In high-stakes scenarios, the bar for explanation quality is higher. It is not enough to show a heatmap of token importances or a generic statement like “the model thinks these keywords are important”—stakeholders need specific, accurate, and understandable justifications. Many existing LLM explainability techniques fall short. Asking an LLM to explain its own output can produce fluent answers, but these self-explanations might themselves be hallucinated or aligned to user expectations rather than the truth. Feature attribution methods can tell which words in the prompt had the most influence on a given token of output, but they don’t necessarily convey why those words influenced the decision in a causal sense. Moreover, evaluating explanations is challenging: an explanation in a regulated context must not only be plausible, but faithful to the model’s actual decision process and acceptable to domain experts. Currently, there is a lack of standardized metrics to evaluate whether an LLM’s explanation is correct and useful, especially when human lives or rights are impacted.
Emerging Approaches for Explainability and Trust in LLMs
Researchers and practitioners are actively exploring new techniques to make LLM outputs more explainable and trustworthy, often drawing on principles from knowledge representation, causality, and rigorous reasoning frameworks.
Evidence-bound reasoning and Retrieval-Augmented Generation (RAG) offer a powerful approach to reduce hallucinations and increase transparency by binding outputs to verifiable evidence [13]. Rather than relying purely on the LLM’s implicit knowledge, the system is designed to retrieve before reasoning and validate before ranking [14,15]. In RAG, the LLM is augmented with an external knowledge base: given a query, the system first fetches relevant documents and then prompts the LLM to generate an answer grounded in those documents, often with explicit citations. This ensures that every factual claim the model makes can be traced back to a source, thereby increasing trust. In enterprise settings, multi-layered AI stacks combine pre-processing, evidence retrieval, logical rules or risk models, and finally an LLM layer to translate structured conclusions into natural language [16,14]. Evidence-bound frameworks typically enforce consistency checks (e.g., cross-validating facts, abstaining when information is insufficient) to prevent unsupported generation [13]. The net effect is improved reliability and more repeatable outputs suitable for enterprise scale [14].
Causal reasoning and graphical models are another promising direction. Traditional LLMs lack an explicit understanding of cause-and-effect; they operate statistically, which can lead to spurious reasoning. In contrast, causal models (such as Bayesian networks or structural causal models) encode relationships between variables and answer “what if” questions through counterfactuals. Researchers propose using causal graphs alongside LLMs to improve explainability [17,18]. In a complex decision support system, a causal graph could model domain factors and dependencies (e.g., medical conditions and symptoms, or microservices and their dependencies) [19]. The LLM can interface with this graph to ensure its explanations respect known causal structures. This enables counterfactual explanations, such as “If the applicant’s income were $10,000 higher, the loan would likely be approved,” which are intuitive and useful in regulated contexts [20]. Causal structures also support root-cause analysis: LLMs may confuse correlated symptoms with causes, while a causal model can isolate likely causes and explain why they produce observed effects [21,18]. Bayesian inference over causal graphs can additionally quantify uncertainty about causes or predictions [22].
Self-reflection and traceable reasoning paths attempt to improve transparency and reliability by making intermediate reasoning explicit. Techniques like chain-of-thought prompting can expose reasoning steps, while newer approaches emphasize self-auditing: before finalizing an answer, the model reviews its own intermediate reasoning for errors or inconsistencies [23]. A self-reflective model might generate a draft solution and rationale, then verify each step for coherence, revising or consulting external tools to improve accuracy [24,25]. This creates a traceable path that can be logged and audited. Empirical results suggest self-reflection can catch contradictions and reduce harmful or biased outputs [26]. Tree-of-Thoughts methods explore multiple reasoning branches and select the most consistent path, offering additional transparency about alternatives considered [27].
Uncertainty estimation and calibration address the trustworthiness question of how confident the model is. LLMs do not naturally provide calibrated uncertainty, so research adapts probabilistic techniques such as ensembles and sampling to estimate output variance. Enterprise deployments can implement uncertainty thresholds: if confidence is below a threshold, the system abstains, triggers fallback workflows, escalates to humans, or flags outputs as low confidence [28,29]. Self-reflective work also explores token-level and sequence-level certainty estimation, allowing the model to detect shaky segments and double-check facts before finalizing [30,25]. These mechanisms reduce overconfident errors—confidently delivered but incorrect outputs [31]—and enable risk-aware workflows where decision-makers can interpret responses alongside confidence estimates [28].
Counterfactual and contrastive explanations extend classic XAI ideas to LLM settings. In many enterprise scenarios, an LLM effectively makes a decision or classification (e.g., fraud tagging, routing). Counterfactual explanations describe minimal changes that would alter the decision, such as “Had the claim amount been under $5000, the claim would have been approved.” Researchers show how LLMs can help generate natural language explanations from sets of counterfactual examples [32]. A pipeline may generate diverse counterfactuals (e.g., with DiCE for structured data or prompt perturbations for text) and then use an LLM to synthesize an explanation in plain language [33,34]. Contrastive explanations clarify why outcome A was chosen over B, which can support policy analysis and regulatory reasoning.
Overall, the frontier of explainable LLM research increasingly favors hybrid systems that augment LLMs with structured reasoning, explicit knowledge, and meta-cognitive checks. The goal is to move toward systems that are not only powerful but also transparent, verifiable, and aligned with institutional expectations of accountability.
Regulatory and Policy Context
The push for explainable and trustworthy AI is not only technical but also regulatory. Policymakers are introducing frameworks that make transparency and explainability core requirements for certain systems. Here, we focus on the EU AI Act and related instruments.
The EU AI Act, expected to take effect in 2026, adopts a risk-based approach and imposes stringent requirements on high-risk AI systems. Many enterprise and government uses of LLMs could fall into high-risk categories (e.g., credit scoring, employment decisions, public service delivery). For high-risk systems, the Act mandates transparency, traceability, and human oversight. Article 30 requires providers to maintain detailed documentation—effectively audit trails—recording decision processes, data, and logic behind outcomes [35,36]. This implies that LLM deployments in high-risk contexts must support reconstruction and explanation of outputs. The Act calls for “appropriate transparency and provision of information to users,” requiring explanations in understandable terms [2]. While general-purpose LLMs are not automatically high-risk, using them within high-risk applications shifts compliance responsibilities to deployers. This drives the need for built-in logging, model version tracking, and explainable outputs to avoid penalties and liability [36,2].
GDPR has also shaped expectations through provisions on automated decision-making. While the phrase “right to explanation” is debated, GDPR requires providing meaningful information about the logic involved when decisions are made solely by automated means with significant effects. This pushes organizations to produce human-centric explanations, not just technical transparency. As noted in industry discussions, tracing reasoning behind predictions is increasingly a must-have for enterprise AI [37].
Beyond general regulations, sectoral standards reinforce explainability. Finance requires strong model risk management (e.g., Basel III expectations). Healthcare and medical devices demand accountability and transparency to prevent patient harm. The U.S. NIST AI Risk Management Framework, though voluntary, emphasizes documentation, performance monitoring, and continuous oversight [40]. Governments also consider algorithmic auditing mandates. These trends converge on auditability as a central requirement; organizations without comprehensive audit infrastructure face operational and legal risks [38], while those investing early benefit from faster approvals and trust [39].
Enterprise AI governance further amplifies these pressures. Organizations increasingly require that decisions affecting customers or citizens be explainable on request, and that AI outputs be logged sufficiently to reproduce decisions. For LLMs, governance may include storing prompts, retrieved documents, chain-of-thought traces (if used), and surrounding context. Governance bodies often adopt model cards or AI fact sheets to communicate intended use and limitations. As one view summarizes: “you can’t govern what you can’t understand,” positioning explainability as the foundation for effective guardrails [41].
Design Proposals for Explainable and Trustworthy LLM Systems
To operationalize these ideas, we propose design and implementation strategies that prioritize transparency without sacrificing performance.
A modular pipeline architecture is often preferable to a monolithic end-to-end LLM. The system can be decomposed into input processing, retrieval, reasoning, generation, and explanation modules, each with explicit logging. This aligns with the “System 1 vs System 2” framing [42], where the LLM provides fluent generation while surrounding modules enforce consistency, factual accuracy, and traceability. Such layered pipelines have been illustrated in security risk contexts, enabling systems that explain decisions and integrate evidence in auditable ways [16]. Modular designs improve transparency, isolate errors through validation stages, and give domain experts access to intermediate outputs that are more interpretable than raw model internals.
Explanation-guided training and fine-tuning aim to make models easier to explain. Rationale-augmented training teaches models to pair answers with justifications, while verification constraints require explanations to be supported by retrievable evidence. Contrastive fine-tuning can teach models to prefer faithful explanations over plausible but incorrect ones. Reinforcement learning can incorporate rewards for explanation quality. Although nascent, these approaches can yield models that produce more structured, inspectable outputs and avoid skipping critical reasoning steps.
Interactive explanation interfaces recognize that explanation needs vary by user role. After an answer, users should be able to ask follow-up questions such as “Why?” “What evidence supports this?” or “What if conditions change?” Systems can use stored traces and evidence to respond. In regulated environments, such interfaces support rights to information and improve usability. Counterfactual queries can trigger re-runs under modified inputs, producing actionable explanations. These interfaces require rich trace storage and a consistent explanation mode that translates logs into accessible narratives.
Confidence modeling should be built-in. Systems can estimate uncertainty via sampling or prompt ensembles and attach confidence labels to outputs. Low-confidence outputs should trigger fallback workflows (human review, abstention, disclaimers). Proper communication of confidence prevents misuse and enhances trust. Logging confidence and fallbacks strengthens the audit trail and demonstrates responsible system design.
Robust audit trails require explicit schemas capturing timestamps, transaction IDs, model versions, prompts, retrieved sources, reasoning traces, confidence scores, and user feedback [43,36]. Proposed “minimum viable audit schemas” emphasize prompt versioning, source document identifiers, and regulatory context tags [44,36]. Such logs enable forensic reconstruction and after-the-fact explainability, ensuring AI decisions are first-class citizens in enterprise audit ecosystems [35].
Finally, human-in-the-loop governance should be integrated. Random sampling, risk-triggered review, and periodic explanation audits enable continuous validation of system behavior. Human reviewers can confirm or correct rationales, feeding back into improvements or preserving annotations for future cases. This operationalizes trustworthiness as an organizational practice, not just a model property.
Discussion and Future Directions
Explainability and trustworthiness remain evolving frontiers. A central open problem is evaluating explanation faithfulness: an explanation must reflect true decision drivers, not merely sound plausible. Standardized metrics are limited, motivating benchmarks that evaluate both answer quality and explanation utility, possibly domain-specific. User-centered evaluation is also necessary, since explanations must support real human decisions.
A persistent tension is the trade-off between model complexity and interpretability [45,12]. Distillation into interpretable surrogates, local approximations, and compositional interpretability are promising directions [46], but scaling them to LLM richness remains difficult. Another tension is between transparency and privacy/security: fully traceable explanations may leak sensitive data or enable system gaming. Tiered explanation access and privacy-preserving explanation methods are underexplored and likely require coordinated policy and technical solutions.
Continual updates further complicate explanation consistency. As models and prompts evolve, explanation interfaces must remain stable and comparable over time. This suggests a future where MLOps expands to “XAI Ops,” with explicit controls ensuring explanation quality and stability across deployments.
Causal and ethical reasoning integration is also promising. Systems that explicitly check causal constraints or policy rules could provide inherently defensible explanations (“I did not recommend X because it violates rule Y”), aligning outputs with governance. Finally, standardization is likely: shared schemas for explanation reporting, audit logs, and model documentation could enable certification and independent audits, analogous to standards in finance and compliance.
Conclusion
Explaining LLM behavior and ensuring trustworthiness is a grand challenge spanning research, industry practice, and policy. The opacity of LLMs creates risks in high-stakes deployments where accountability, auditability, and legal defensibility are required. Emerging approaches—evidence-bound reasoning, causal structures, self-reflection, and uncertainty estimation—offer promising paths. Regulatory frameworks, particularly the EU AI Act, are accelerating the shift from optional transparency to mandatory explainability.
An explainable and trustworthy LLM system is one that can trace its reasoning, justify outputs with evidence, quantify uncertainty, and support human oversight. Achieving this requires hybrid architectures, rigorous evaluation, and operational governance. As regulation and adoption expand, explainability becomes central to sustainable deployment rather than a peripheral feature.
References
[1] Ahsan Bilal et al. “LLMs for Explainable AI: A Comprehensive Survey.” (2025). https://arxiv.org/html/2504.00125v1
[2] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[3] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[4] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[5] “A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models.” (2024). https://arxiv.org/abs/2407.02646
[6] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[7] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[8] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[9] Ahsan Bilal et al. “LLMs for Explainable AI: A Comprehensive Survey.” (2025). https://arxiv.org/html/2504.00125v1
[10] Ahsan Bilal et al. “LLMs for Explainable AI: A Comprehensive Survey.” (2025). https://arxiv.org/html/2504.00125v1
[11] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[12] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[13] “When ‘Good Enough’ Hallucination Rates Aren’t Good Enough...” Balbix (2023). https://www.balbix.com/blog/hallucinations-agentic-hype/
[14] “When ‘Good Enough’ Hallucination Rates Aren’t Good Enough...” Balbix (2023). https://www.balbix.com/blog/hallucinations-agentic-hype/
[15] “When ‘Good Enough’ Hallucination Rates Aren’t Good Enough...” Balbix (2023). https://www.balbix.com/blog/hallucinations-agentic-hype/
[16] “When ‘Good Enough’ Hallucination Rates Aren’t Good Enough...” Balbix (2023). https://www.balbix.com/blog/hallucinations-agentic-hype/
[17] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[18] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[19] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[20] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[21] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[22] “How Causal Reasoning Addresses the Limitations of LLMs in Observability.” Causely (2025). https://www.causely.ai/blog/how-causal-reasoning-addresses-the-limitations-of-llms-in-observability
[23] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[24] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[25] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[26] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[27] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[28] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[29] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[30] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[31] “Self Reflection and Fixing Inconsistency in Language Models.” Galileo (2025). https://galileo.ai/blog/self-reflection-in-language-models
[32] “LLMs for Explainable AI: A Comprehensive Survey.” (2025). https://arxiv.org/html/2504.00125v1
[33] “Using LLMs for Explaining Sets of Counterfactual Examples to Final Users.” (2024). https://arxiv.org/html/2408.15133v1
[34] “Using LLMs for Explaining Sets of Counterfactual Examples to Final Users.” (2024). https://arxiv.org/html/2408.15133v1
[35] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[36] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[37] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[38] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[39] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[40] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[41] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[42] “When ‘Good Enough’ Hallucination Rates Aren’t Good Enough...” Balbix (2023). https://www.balbix.com/blog/hallucinations-agentic-hype/
[43] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[44] Lawrence Emenike. “Audit Trails and Explainability for Compliance...” Medium (Dec 2025). https://lawrence-emenike.medium.com/audit-trails-and-explainability-for-compliance-building-the-transparency-layer-financial-services-d24961bad987
[45] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability
[46] “LLMs for Explainable AI: A Comprehensive Survey.” (2025). https://arxiv.org/html/2504.00125v1
[47] “LLM explainability: Trust and transparency in gen AI.” GoML (2024). https://www.goml.io/blog/llm-explainability