By Alan Shark
This article was originally published by Route Fifty Republished here with the author’s permission.
Artificial intelligence systems are now making decisions in policing, hiring, healthcare, cybersecurity, purchasing and finance — but errors or biases can have significant consequences.
Humans alone can’t keep up: models are too complex, too fast, too large in scope. And yet, nearly every AI policy states humans must provide oversight and control. Keeping up with advancements in AI applications is almost impossible for humans. Worse, some admit to over-reliance on AI applications. This is where the idea of AI systems designed to check other AI systems comes in.
Traditionally, humans have performed this role. Auditors, compliance officers, regulators and watchdog organizations have long worked to ensure systems operate as intended. But when it comes to AI, humans alone may no longer be enough. The models are too complex, too fast, and too embedded in decision-making pipelines for manual oversight to keep pace.
That’s why researchers and practitioners are turning to an intriguing solution: using AI itself to audit AI. Recognizing the impact of AI on government applications, in 2021, the Government Accountability Office developed an ahead-of-its-time report, “Artificial Intelligence — An Accountability Framework for Federal Agencies and Other Entities.” Although the framework was practical and far-reaching, it still relied on human planning and oversight.
Today, we are entering a new area of AI accountability with talk about the advent of “watchdog AIs” or “AI auditors” that test, verify and monitor other AI models. This is increasingly important as AI grows more complex and less transparent to human reviewers.
Making the case for AI auditing, we can safely assume that AI can rapidly analyze outputs across millions of data points. And unlike human auditors, AI doesn’t get tired or overlook details. Auditing can occur in real-time, and flag problems as they arise. AI auditors can probe “black box” models with tests humans couldn’t do manually. Taken together, AI auditing strengths can be summarized by its ability to scale, provide consistency, speed, transparency, and accuracy.
Auditing AI is not a single technology but a suite of methods. Some of the most promising approaches include:
- Adversarial testing: One AI generates tricky edge cases designed to trip up another AI, exposing blind spots.
- Bias and fairness detection: Auditing systems measure outcomes across demographic groups to reveal disparities.
- Explainability tools: Specialized models analyze which factors most influenced a decision, helping humans understand why a model reached its conclusion.
- Continuous monitoring: AI can watch for “model drift” — when performance degrades over time as data or circumstances change — and signal when retraining is needed.
In many ways, this mirrors how cybersecurity works today, where red teams and intrusion-detection systems constantly test defenses. Here, the target is not a firewall but another algorithm.
Real-world applications are emerging, though still in its early stages, AI auditing is moving beyond theory. Here are several examples:
- Finance: Some firms are already deploying AI to double-check fraud-detection models, ensuring that suspicious activity flags are consistent and not biased.
- Healthcare: AI-driven validation tools are being used to test diagnostic algorithms, checking their accuracy against known patient outcomes.
- Cybersecurity: “Red team” AIs are being trained to attack models the way hackers might, helping developers harden systems before release.
- Public sector pilots: Governments are beginning to experiment with algorithmic auditing programs, often in regulatory “sandboxes” where new models are tested under close supervision
These examples suggest a growing recognition that human oversight must be paired with automated oversight if AI is to be trusted at scale. At the same time, we must acknowledge AI auditing risks and limitations raise their own set of challenges. This includes the following:
- The infinite regress problem: If one AI audits another, who audits the auditor? At some point, humans must remain in the loop. Or perhaps there might be a third level of AI checking on AI, checking on AI.
- Shared blind spots: If both models are trained on similar data, they may replicate the same biases rather than uncover them.
- Over-trust: Policymakers and managers may be tempted to rely too heavily on “AI-certified AI” without questioning the underlying process.
- Resource costs: Running parallel AI systems can be expensive in terms of computing power and energy consumption.
In short, as tempting as it may appear, AI auditors are not a panacea. They are tools—powerful ones, but only as good as their design and implementation.
This raises critical governance questions. Who sets the standards for AI auditors? Governments, industry consortia, or independent third parties? Should auditing AIs be open-source, to build public trust, or proprietary, to protect against exploitation? And how do we ensure accountability when the auditors themselves may be opaque? Can or should AI auditing be certified, and if so, by whom?
There are strong arguments for third-party, independent auditing — similar to how financial auditing works today. Just as markets rely on trusted external auditors, the AI ecosystem will need its own class of independent algorithmic auditors. Without them, self-auditing could resemble letting the fox guard the henhouse.
Most experts envision a layered approach: humans define auditing standards and interpret results, while AI handles the heavy lifting of large-scale checking. This would create multiple levels of defense — primary AI, auditing AI and human oversight.
The likely result will be a new industry built around AI assurance, certification, and compliance. Just as accounting gave rise to auditing firms, AI may soon give rise to an “AI auditing sector” tasked with keeping digital systems honest. And beyond the technical details lies something more important: public trust. The willingness of people to accept AI in critical domains may depend on whether robust and credible audit mechanisms exist.
AI auditing AI may sound strange at first, like machines policing themselves. But far from being a case of “the fox guarding the henhouse,” it may prove essential to making AI safe, reliable and trustworthy. The truth is, humans cannot realistically keep up with the scale and complexity of today’s AI. We need allies in oversight — and in many cases, the best ally may be another AI. Still, human judgment must remain the final arbiter.
Just as financial systems depend on auditors to ensure trust, the AI ecosystem will need its own auditors—both human and machine. The future of responsible AI may well depend on how well we design these meta-systems to keep each other in check.
Dr. Alan R. Shark is a senior fellow at the Center for Digital Government and an associate professor at the Schar School for Policy and Government, George Mason University, where he also serves as a faculty member at the Center for Human AI Innovation in Society (CHAIS). Shark is also a senior fellow and former Executive Director of the Public Technology Institute (PTI). He is a Fellow of the National Academy of Public Administration and Founder and Co-Chair of the Standing Panel on Technology Leadership. Shark is the host of the podcast Sharkbytes.net.