The New Meaning of Compression
When we hear “compression,” we usually think of making a file smaller. But in the context of modern AI, it refers to a different kind of shrinking. Advanced AI systems can perform millions of calculations, analyse decades of data, and run countless simulations
in seconds, delivering a single, neat recommendation. This output is a form of compressed communication—an answer that summarises an immense, invisible analytical process. For example, an AI might recommend a complex stock trade or a change to a global supply chain. The recommendation is simple, but the reasoning behind it is a black box of interconnected variables that no human could track in real time. This creates what experts call the “oversight paradox”: the more work we delegate to a highly capable AI, the less we understand the work ourselves. Human oversight, intended as a critical safeguard, is at risk of becoming a hollow gesture.
From Safeguard to Signature
The primary risk of this new compression is that human oversight degrades into a simple rubber stamp. This happens through two powerful forces. The first is automation bias. Studies have shown that when a system is correct most of the time, humans naturally start to trust it blindly. A manager reviewing AI-generated reports will, over time, stop questioning the details and start simply approving the output. They drift from actively evaluating to passively confirming. The second, more dangerous force is deskilling. The core competence needed to oversee an AI system is built through practice. A radiologist who expertly spots anomalies does so because they have reviewed thousands of scans. But if an AI now does the primary review, the human expert gets fewer and fewer opportunities to exercise their own judgment. Their hard-won expertise begins to fade. When the overseer can no longer perform the task without the AI, their ability to spot a subtle error or a biased conclusion disappears. At that point, their approval is no longer a safeguard; it is merely a signature.
The Key Qualification: Applied Interpretability
The key qualification needed to counter this trend is not about learning to code or build AI models. Instead, it is a human-centric skill we can call “applied interpretability.” This is the active ability to make an AI explain its reasoning in human-understandable terms and to critically evaluate that explanation. Interpretability is about demanding transparency from our powerful but opaque AI tools. It’s a competency built on three pillars. First is deep domain expertise—the overseer must remain an expert in their field, independent of the AI. Second is a healthy dose of critical thinking, which involves questioning the AI's outputs, asking what data it used, and probing for potential biases. Third is a functional understanding of AI capabilities and limitations, knowing where it excels and where its blind spots are. An employee with applied interpretability doesn't just accept the AI’s recommendation. They ask, “Why did you conclude this? What factors did you weigh most heavily? What scenarios did you ignore?” This transforms them from a passive user into an active and essential partner in the decision-making process.
How to Cultivate the Human Safeguard
Developing this qualification requires a conscious effort from both individuals and organisations. For professionals, it means resisting the urge to passively accept AI-driven answers. It involves staying engaged with the fundamental tasks of their profession and treating AI as a collaborator to be questioned, not an oracle to be obeyed. This requires a commitment to continuous learning in one's own field. For organisations, the solution is threefold. First, invest in training that focuses on data literacy, critical thinking, and ethical reasoning, not just on how to use a new software tool. Second, build governance structures that demand transparency. This could mean using AI systems that are designed for explainability or establishing review processes that require humans to document their validation of an AI's decision. Finally, leaders must actively protect the skills of their human experts. This might involve rotating tasks, ensuring people still perform core work manually on a regular basis, and rewarding employees who successfully challenge and improve upon AI-generated insights. The goal is to keep the human in the loop not just for show, but as a truly capable and engaged mind.
