Explainability Is the Measure of Accountability: Why Responsible AI Must Evolve into Accountable AI

As I was wrapping up discussions with some people that I had trained for the Project Management Professional (PMP) certification, discussions emerged around Responsible AI. It was interesting to read people's thought about "Responsible AI" from a technical and career enablement standpoint! But, Responsible AI is the start of the leadership for holding oneself accountable for the decisions made. I think therefore, there needs to be an evolution of the next level of AI leadership foundations! 

Undoubtedly, Artificial Intelligence has become remarkably proficient at making predictions. It identifies fraudulent transactions, recommends medical diagnoses, screens job applicants, prioritizes customer requests, and increasingly influences decisions once reserved for human judgment. Yet as organizations race to adopt AI, they continue asking the wrong question. The question is not whether AI is intelligent enough to make decisions. The real question is whether leaders are accountable for the decisions AI helps make. Responsible AI has advanced the conversation around technical reasoning but still has not advanced the ethical leadership. 

It is important to understand that responsibility without accountability remains an aspiration rather than an operational reality. We need to shift our mindset towards the impact of our actions (or lack of it) on the people (employees, customers, end users, societies) and planet (environment) by asking, "How are we accountable for the consequences of an AI-enabled decision?" This mindset is where I believe we move beyond Responsible AI toward Accountable AI. 

• Responsible AI establishes the intention to do the right thing. 
• Accountable AI requires us to explain, justify, and ultimately own what happens.

Consequently, the future belongs to Accountable AI, where both explainability and explicability becomes the mechanism through which accountability is demonstrated.

"Accuracy measures performance. Explainability measures reasoning behind the prediction. Explicability measures ethical and social accountability"
— Dr. Shri Rajagopalan

The distinction begins with three concepts that are often treated as interchangeable: accuracy, explainability, and explicability. 

• Accuracy tells us whether the prediction was correct. 
• Explainability helps us understand the reasoning behind the prediction by understanding data, variables, patterns, features or user behaviors contributed to the model's conclusion. This is essential because an organization cannot govern what it cannot understand. 
• But explainability has a boundary. Knowing why an algorithm reached a conclusion does not tell us whether the conclusion should have been reached, whether the underlying assumptions were appropriate, or whether acting upon that conclusion is ethically defensible. That requires explicability (Floridi & Cowls, 2019). 

Leaders are not faced with explaining the technical architecture or the model differences in the AI system used. That would be explainability based on accuracy and precision. They are faced with answering beyond technical reasoning focused on the impact of the AI system. Hence, explicability takes us beyond technical reasoning toward the ethical and social accountability of AI. It asks not only “Why did the system make this prediction?” but also “Should we accept this prediction, what are its consequences, and can we justify those consequences to the people and communities affected?”

Consider an AI system used to screen employment applicants. The model may be highly accurate and completely explainable. We might even be able to identify precisely which characteristics contributed to a candidate's rejection. But what if those characteristics reflect historical patterns of discrimination embedded within the training data, such as what happened with the Amazon's AI recruiting tool showing preferential treatment to men over women candidates (Datin, 2018)? Explainability can tell us why the model rejected the candidate. Explicability asks the more difficult question: Should those factors have been permitted to influence the decision at all? This is where algorithmic bias becomes a leadership challenge rather than simply a data science problem. The issue is not merely whether the algorithm works as designed. It is whether the organization designed, deployed, and governed the system in a way that is ethically and socially defensible.

Every participant in the AI value delivery pipeline therefore has a responsibility to ask questions that technology alone cannot answer. The onus now shifts to everyone to think risk based questions on outcome and value! Not just, "Can we get this faster?" but more focused on:

• Who benefits from this technology? 
• Who bears its risks? 
• Who controls the algorithmic system? 
• Whose voices are suppressed? 
• Which communities are marginalized? 
• What meaningful options do we give the users affected by the system? 
• What unintended consequences might emerge over time? 

These questions move us from technical accountability toward societal accountability. They also recognize that AI does not operate in a vacuum. An algorithm can optimize a decision while simultaneously optimizing inequality. It can improve overall performance while imposing disproportionate harm on a particular community. It can be technically correct and ethically wrong. That is why the ethical challenge is not the technology itself but the leadership's governance of its use.

Explicability also gives us a different way to understand the fundamental ethical principles of beneficence, non-maleficence, justice, autonomy, and fidelity. These principles should not be treated as independent boxes on an AI ethics checklist. They are symbiotic conditions of accountability. 

• If an AI system causes preventable harm, non-maleficence has been compromised. 
• But the consequences rarely stop there. If the affected individuals have no meaningful ability to challenge the decision, autonomy is compromised. 
• If the harm disproportionately affects a particular population, justice is compromised. 
• If the technology's promised benefits are not realized or are realized primarily by those already advantaged, beneficence is compromised. 
• And when an organization cannot explain, justify, or take ownership of these consequences, fidelity and trust begin to disappear. 

When one ethical principle is compromised, the others are placed at risk requiring not only understanding the primary risk but also the secondary risks. Ethical accountability therefore cannot be achieved by satisfying five principles independently; it requires recognizing their interdependence. The need for this awareness is why I propose the movement from Responsible AI to Accountable AI (Rajagopalan, 2020). If an end-user was unfairly treated with no alternatives presented, it is not technology failure but leadership failure. Reasoning at that point is too late as the arrow has not only left the bow but hit the wrong person or the wrong place. Explainability, from Responsible AI, provides evidence of reasoning. Explicability, from Accountable AI, provides the basis for ethical judgment. This distinction is critical because a transparent explanation of an unjust decision does not make the decision just. A perfectly explainable harmful outcome is still harmful.

Ultimately, Accountable AI requires humans to remain meaningfully involve, not simply as a Human-in-the-Loop who approves or rejects an algorithmic recommendation, but as Humans-in-the-Middle of an ongoing accountability system. Anyone in the AI based delivery lifecycle shouldn't be afraid of their jobs but should reorient themselves to their new improved leadership responsibility. The future of ethical AI, therefore, should not be measured solely by how intelligent our machines become. It should be measured by how accountable the humans governing those machines remain in our commitment to people, community, and the planet.

“Responsible AI asks whether we are doing the right things. Accountable AI asks whether we are willing to stand behind what those things do.”
— Dr. Shri Rajagopalan

What are your thoughts?

References

Datin, J. (2018). Insight - Amazon scraps secret AI recruiting tool that showed bias against women. https://www.reuters.com/article/world/insight-amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK0AG/

Floridi, L., & Cowls, J. (2019). A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review, 1(1). https://doi.org/10.1162/99608f92.8cd550d1

Rajagopalan, S. (2020, September). Artificial Intelligence Solutions: Four Considerations extended from Digital Bioethics. https://agilesriram.blogspot.com/2020/09/artificial-intelligence-solutions-four.html