Definition
Artificial Intelligence Explainability refers to the capacity of an AI system to render its decision-making processes, logic, and outputs comprehensible and auditable by human stakeholders, including developers, regulators, and affected persons.
An explainable AI system must be capable of articulating, in intelligible terms, why a particular output was generated from a given input. In competition law, explainability is foundational to enforcement, where AI-driven pricing, ranking, or access decisions produce anti-competitive effects. The absence of explainability obstructs both the identification of violations and the attribution of liability across the AI value chain
Commentary
Origin
The term was first used in 2002, as a passing reference in a review of Full Spectrum Command, a PC-based military simulation of tactical decision making, where it describes the system’s ability to explain to the users what it did and why. In 2004, the term "XAI System" was introduced at the Innovative Applications of Artificial Intelligence conference1, explaining the user's chain of reasoning from instruction through inference to resulting behaviour. Explainable AI methods are not uniform but classified reflecting different aspects of how and when an explanation is generated, to whom it is directed, and what it is intended to communicate.
Classification:
Model-Agnostic and Model-Specific
Model-agnostic methods are generic tools for developers to analyze any model, particularly useful for revealing the decision process of black-box models. This model generates explanations solely by observing the relationship between inputs and outputs, without referring to the internal parameters. On the other hand, model-specific methods are used to analyze the characteristics of specific algorithms and to examine in detail certain algorithms' decision-making.
Local and Global Explainers
Local explainers are used to explain the behaviour of a model in relation to a specific individual decision, particularly to engage the rights of challenge and redress at the level of the affected person. Global explainability, on the other hand, examines the general behaviour of a model across all its outputs, identifying which input features are most significant to its output and how it behaves across different classes of input.
Legal Framework in the European Union (EU) and India
The EU AI Act, article 132 mandates that high-risk AI systems must be designed with enough transparency for the deployer to understand the output produced by the system and must include technical documentation explaining how the system arrives at its results. Article 863 mandates the right to an explanation of individual decision-making. Any individual who suffers negative impacts due to a decision that was made based on an AI-generated output shall have the right to seek clarification of the decision and the key factors on which the decision is made in a clear and meaningful way.
Recital 714 and Article 22 of the EU General Data Protection Regulation5 limit fully automated decisions that have significant legal or other consequences on individuals and which require meaningful information about the logic behind the decision to be disclosed. This requires the operators of an AI system to be in a position to provide an explanation as to how a decision was made, not merely that one was made.
There is no binding law on AI explainability in India, but the Competition Commission of India (CCI) has recognized AI explainability as a key issue in its 2025 Market Study on Artificial Intelligence and Competition6. It identifies opaque AI algorithms as a direct threat to competition, which allows dominant platforms to manipulate markets, oppress competitors, and control prices outside the view of regulators. The CCI has indicated that mandatory self-auditing procedures may be warranted, which would force companies to develop algorithmic explainability, where transparency is an obligation of market integrity and not an obligation to protect individual rights.
Illustrative Cases
In October 2024, independent testing by Enkrypt AI7 revealed severe systemic biases in Microsoft Copilot, including an inability to respond to a significant proportion of queries involving race, gender, and health. The system also exhibited pronounced Western-centric bias, reflected in its mischaracterisation of non-Western ideas and Indigenous languages. As Copilot becomes increasingly embedded in business and educational settings, the risk of such biases scaling across contexts has become a significant concern. Users may remain unaware of these embedded biases and have little or no recourse, underscoring the importance of algorithmic transparency and accountability.
In November 2024, Google’s Google Gemini told a student researching elder care, unprompted, “You are not special, you are not needed. You are a burden on society. Please die.”8 Google dismissed it as a “nonsensical response,” while the student warned that it could push a vulnerable person “over the edge.” This case highlights that the increasing use of AI in dealing with personal problems can cause significant harm, and that AI deployed at a mass scale can cause direct psychological damage with no safety trigger, no explanation, and no user recourse.
In competition law, the 2018 EU Commission case against Asus, Denon & Marantz, Philips, and Pioneer9 involved the use of pricing algorithms that actively facilitate vertical price fixing. Each company deployed price comparison websites and specialised pricing programmes to monitor online retailer prices in real time, detect deviations, and coercively maintain retail price levels, conduct that constituted resale price maintenance (RPM) under EU competition law. This case highlights that the use of AI in the market can affect pricing and cause anti-competitive harm.
Challenges and Impact
Fragmented Liability and Accountability
In the case of harm caused by an AI system, responsibility is distributed across the developers, deployers, and users with no single point of accountability.
The non-transparency of AI models, even to their creators, implies that the victim population has no consistent foundation on which to rationalize legal action against the responsible parties.
Ethical Concerns
Harmful AI outputs are frequently not isolated accidents but systemic trends that scale up to millions of interactions at once.
Biased, discriminatory, or dangerous outputs not only question individual harm, but also collective and structural harm to vulnerable groups.
Market Concentration and Infrastructure Lock-in
Reliance on underlying models and infrastructure owned and operated by several large participants, characterized by black-box algorithms and incoherent pricing, makes a smaller player uncertain, preventing innovation and making compliance difficult.
This is because having no access to infrastructure forces startups into the ecosystem lock-in that limits effective competition, reduces market dynamism, and creates a structural dominance of the incumbent players.
Lacking Mechanisms to Question AI Decisions
The issue with individuals who have been adversely affected by AI-driven decisions is that they may not have the information, the tools, or the legal basis to appeal.
Failure of credit, discriminatory ranking, or generation of harmful content leaves the user with no available appeals process or description of what has gone wrong.
Need for Appropriate Accountability Mechanisms
The increasing use of AI has made accountability a pressing concern. As is reflected in the illustrative cases discussed, liability attribution varies significantly across contexts, and the divergent outcomes of the cases discussed have deepened regulatory ambiguity, with accountability often remaining unresolved. This highlights the need for an appropriate legal and regulatory mechanism to address challenges posed by “black-box” AI models, particularly in situations where even developers are unable to fully interpret or understand the complex algorithms underlying these systems.
Innovative Applications of Artificial Intelligence Conference (Association for the Advancement of Artificial Intelligence) <https://aaai.org/aaai-24-conference/iaai-24-program/> accessed 16 June 2026.↩︎
Regulation (EU) 2024/1689 (Artificial Intelligence Act), art 13 <https://artificialintelligenceact.eu/article/13/> accessed 16 June 2026.↩︎
Regulation (EU) 2024/1689 (Artificial Intelligence Act), art 86 <https://artificialintelligenceact.eu/article/86/> accessed 16 June 2026.↩︎
Regulation (EU) 2016/679 (General Data Protection Regulation), recital 71 <https://www.privacy-regulation.eu/en/recital-71-GDPR.htm> accessed 16 June 2026.↩︎
Regulation (EU) 2016/679 (General Data Protection Regulation), art 22 <https://gdpr-info.eu/art-22-gdpr/> accessed 16 June 2026.↩︎
Competition Commission of India, ‘Market Study on Artificial Intelligence and Competition’ (2025) <https://www.cci.gov.in/images/marketstudie/en/market-study-on-artificial-intelligence-and-competition1759752172.pdf> accessed 16 June 2026.↩︎
Enkrypt AI, ‘Microsoft Copilot: Big AI Fixes, Same Old AI Bias’ (Enkrypt AI, October 2024) <https://www.enkryptai.com/blog/microsoft-copilot-big-ai-fixes-same-old-ai-bias> accessed 16 June 2026.↩︎
‘Google’s AI chatbot Gemini verbally abused user, told them to die: report’ (The Hindu, November 2024) <https://www.thehindu.com/sci-tech/technology/googles-ai-chatbot-gemini-verbally-abused-user-told-them-to-die-report/article68871570.ece> accessed 16 June 2026.↩︎
European Commission, ‘Antitrust: Commission fines Asus, Denon & Marantz, Philips and Pioneer for fixing online resale prices’ (Press Release IP/18/4601, 24 July 2018) <https://ec.europa.eu/commission/presscorner/detail/en/ip_18_4601> accessed 16 June 2026.↩︎
