Introduction: The Limits of “Explainable”

In the rush to make AI systems more trustworthy, “explainability” has become the go-to solution. From loan approvals to predictive policing, technologists and regulators alike often call for explainable models. But explainability is not a substitute for accountability. As legal scholars Selbst and Barocas have argued, technical interpretability doesn’t necessarily illuminate the broader institutional and societal impacts of automated decisions.

We can’t stop at showing how a system works — we must also ask who it serves, who it harms, and who answers when it fails.

What Real AI Accountability Looks Like

Accountability in AI is about power, responsibility, and recourse — not just transparency. It’s a commitment to making systems not only intelligible, but also answerable.

1. Redress Mechanisms

Explainability often tells users what happened. Accountability ensures they have the right to challenge it. Without mechanisms for appeal or redress, affected individuals are left powerless — especially when decisions are made by opaque or unregulated models.

“Accountability is about the ability to enforce consequences — not merely explain intent.”

2. Training Data Transparency

Bias doesn’t just emerge during model execution — it’s embedded in the data itself. Yet most large-scale models obscure where their training data comes from. The authors of the “Datasheets for Datasets” initiative emphasize the importance of disclosing how data was collected, labeled, and cleaned to support ethical AI deployment .

3. Oversight & Governance

We need more than ethics principles. We need institutions — regulatory bodies, ethics boards, third-party audits — that can intervene and enforce when systems cause harm. The EU’s AI Act and the White House Blueprint for an AI Bill of Rights are steps in this direction, aiming to provide operational frameworks for enforceable accountability

Why Explainability Isn’t Enough

While interpretability can help engineers debug systems and users understand decisions, it can also be a false comfort. A model may be technically transparent yet ethically flawed — optimizing for outcomes that reinforce inequality.

In her work, Kate Crawford warns against “transparency theater” — the illusion of clarity without meaningful governance [6].

“Explainability without accountability is like a manual for a machine you can’t turn off.”

Building for the Public Interest

True accountability requires alignment with democratic values and public interest. That means:

• Inclusive design: centering marginalized communities in system development

• Ethical procurement: governments and firms choosing AI vendors based on rigorous standards

• Auditable systems: models that can be interrogated by independent third parties

Only when these elements are embedded from the start can we say a system is built for accountability — not just compliance.

Conclusion: Accountability Is an Infrastructure

If we want AI systems to be fair, trustworthy, and just, we can’t rely on technical clarity alone. We need a full-stack approach: from data collection to deployment, from model behavior to institutional oversight.

Explainability is a feature.
Accountability is a foundation.

And it’s time we build accordingly.

Cited References

1. Selbst, A. D., & Barocas, S. (2018). The Intuitive Appeal of Explainable Machines. Fordham Law Review, 87(3), 1085–1139.

2. Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press.

3. Gebru, T., et al. (2018). Datasheets for Datasets. 

4. European Commission. (2021). Proposal for a Regulation on Artificial Intelligence (AI Act)

5. White House Office of Science and Technology Policy. (2022). Blueprint for an AI Bill of Rights

6. Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.