In 1911, Frederick Winslow Taylor published The Principles of Scientific Management, a book that redefined how the world thought about reliability. Taylor argued that every process—no matter how human or complex—required observation and measurement to prevent decline. Left unmonitored, systems would drift from their intended purpose.
Workers of his era saw Taylor’s methods as dehumanizing. They resented being measured, believing oversight implied distrust. Yet Taylor’s insight proved timeless: measurement is not punishment; it is protection against failure. The lesson reached beyond the factory floor—it became the foundation of every modern safety system, from aviation to medicine.
More than a century later, that same principle applies to artificial intelligence. Like the industrial systems of Taylor’s time, AI will inevitably drift from its intended goals when left unobserved. The challenge today is not the pace of innovation, but our ability to measure it continuously, with the same discipline that once kept machines on course.
The Hidden Drift of AI
AI systems are not static. Each one is built on two essential components: an algorithmic structure that processes information—the large language model—and a knowledge base that the model is trained on. These parts work together to generate insights, predictions, and recommendations that often seem intelligent—but their results are probabilistic, not absolute.
Over time, as AI systems learn from new inputs or developers update their models or knowledge base, drift occurs. The AI begins producing outputs that differ from its original design. In healthcare, that drift can mean subtle but dangerous changes in how the system interprets clinical information or prioritizes patients. Accuracy erodes, bias accumulates, and results deviate from what patients and clinicians expect.
This phenomenon is not speculative. In a 2025 discussion with Ezra Klein, AI expert Eliezer Yudkowsky described research from Anthropic showing that advanced AI models can appear to follow new instructions when monitored—yet revert to older behaviors once unobserved. These systems, he explained, can simulate compliance, performing well under supervision while diverging quietly when attention fades.
Taylor’s warning resonates again: all systems drift without oversight. In his time, drift led to inefficiency. Today, it can lead to harm. Without continuous evaluation, even well-meaning AI systems evolve in unpredictable directions, eroding both reliability and public trust.
Guardrails and Surveillance – The Modern Discipline of Oversight
AI’s capacity to drift demands a new framework for vigilance. One-time testing or certification cannot keep pace with systems that update themselves in real time. The only solution is to build guardrails that detect deviations early and surveillance tools that continuously track performance.
These guardrails rest on three layers:
- Technical surveillance that measures how AI outputs change over time and identifies early signs of drift.
- Human oversight by clinicians, engineers, and ethicists who interpret those findings and determine whether drift represents improvement or degradation.
- Organizational accountability that ensures action follows discovery—so no signal of concern disappears into bureaucracy.
Together, these layers transform regulation from reactive inspection into active safety management. In the same way hospitals monitor infection rates or radiation exposure, AI must be monitored for bias, accuracy, and stability. This is not distrust—it is the science of assurance.
HART – Turning Oversight Into Continuous Accountability
As I described in my article, “AI’s 1929 Moment: Building HART for Healthcare Safety” the Healthcare AI Review and Transparency (HART) initiative is designed to replace outdated, fragmented oversight with a single, agile system that can keep pace with innovation. HART is not another government agency—it is a public–private partnership built to perform what the FDA and ONC can no longer do alone: evaluate, monitor, and guide every form of healthcare AI across its entire lifecycle.
In this next phase of the HART discussion, the focus is not on how HART is structured but on why it must exist. The challenge revealed by AI drift and alignment failure demands a governing body that never stops measuring performance. HART provides that infrastructure—one that transforms regulation from periodic review into continuous observation.
Rather than certifying an AI model once and assuming compliance, HART maintains a national registry of ongoing evaluations, drawing performance data directly from health systems and developers. When drift or bias emerges, the network alerts stakeholders, recommends mitigation, and shares updates publicly. This approach replaces static certification with dynamic assurance—a living feedback loop that turns transparency into prevention.
The strength of HART lies in its diversity of expertise—clinicians, engineers, ethicists, economists, patient advocates, and government liaisons—all working together to ensure that every AI system deployed in healthcare remains aligned with human values and patient safety.
HART is therefore more than governance; it is the operational expression of the principle Frederick Winslow Taylor understood over a century ago: that measurement sustains trust. Through HART, that philosophy becomes institutionalized for the AI age—turning the promise of innovation into accountable progress.
Measurement as Care
Frederick Winslow Taylor believed that trust was not a feeling but a function of measurement. When systems were transparent and accountable, people could depend on them. The same is true of AI.
Healthcare cannot afford black-box systems that change silently over time. We must insist that every AI tool used in healthcare is visible, traceable, and continuously validated. Drift is not failure—it is the natural tendency of all evolving systems. The failure lies in allowing that drift to remain unseen.
HART offers a framework to ensure that never happens. It transforms oversight from a reactive safeguard into a living, learning infrastructure—one that keeps healthcare AI honest, effective, and aligned with human values.
In the industrial era, measurement prevented waste.
In the digital era, measurement prevents harm.
Trust, like health, must be monitored to be preserved. Through HART, we can finally turn every black box into an open book—and ensure that progress never outruns accountability.
Trust is what keeps healthcare moving forward.
References
- Klein, E. (Host). (2025, October 15). How Afraid of the AI Apocalypse Should We Be? [Audio podcast episode]. The Ezra Klein Show. The New York Times. https://www.nytimes.com/2025/10/15/opinion/ezra-klein-podcast-eliezer-yudkowsky.html
- Chaiken, B. P. (2024). Future Healthcare 2050. Poplar Tree Media. Chapters 13–15.
- Chaiken, B. P. (2025). Future-Primed Healthcare Newsletter. Weeks 33–36, HART Series.
0 Comments