The Hidden Cost of AI Efficiency

by

Author: Denis Avetisyan

As artificial intelligence reshapes work, a new study examines how productivity gains can subtly erode human skills and dignity.

This review proposes a framework for building ‘sociotechnical immunity’ to address asymptomatic harms from AI-driven workflow automation and ensure dignified human-AI interaction.

While artificial intelligence promises unprecedented gains in productivity, its subtle erosion of human expertise presents a critical paradox. This paper, ‘From Future of Work to Future of Workers: Addressing Asymptomatic AI Harms for Dignified Human-AI Interaction’, shifts focus from workplace transformation to the wellbeing of workers, revealing how AI can induce ‘intuition rust’ and skill atrophy even as performance initially improves. Through a year-long study of cancer specialists, we demonstrate that these asymptomatic effects evolve into chronic harms like identity commoditization, demanding a new framework for dignified human-AI interaction. Can we proactively build ‘sociotechnical immunity’-mechanisms that safeguard worker expertise and agency alongside institutional goals-to ensure a future where AI truly augments, rather than diminishes, human capabilities?

The Subtle Erosion of Expertise in an Age of AI

The promise of artificial intelligence as a tool to augment human work is increasingly shadowed by the ‘AI-as-Amplifier Paradox’, a phenomenon where reliance on these systems subtly erodes the very skills they were meant to enhance. Rather than fostering a collaborative synergy, consistent delegation of complex tasks to AI can lead to a gradual deskilling of professionals. This isn’t a case of immediate obsolescence, but a slow diminishment of core competencies as individuals become overly dependent on algorithmic solutions. The paradox lies in the fact that while AI improves immediate output and efficiency, it simultaneously diminishes the underlying human capital, potentially creating a future workforce less adaptable and innovative than the one it sought to empower. The long-term consequences of this trend extend beyond individual skillsets, impacting an organization’s resilience and capacity for genuine problem-solving.

The subtle shift in workforce dynamics driven by artificial intelligence isn’t manifesting as widespread job losses, but rather as a gradual decline in fundamental human capabilities, a phenomenon termed ‘Asymptomatic Effects’. Studies suggest that over-reliance on AI tools for tasks previously demanding skill and judgment leads to a measurable erosion of those very skills, even while apparent productivity remains high. This isn’t a catastrophic failure of competence, but a creeping deskilling, where professionals become increasingly dependent on algorithmic assistance and less able to perform independently. The concern isn’t that AI is immediately taking jobs, but that it’s quietly diminishing the core expertise that underpins entire professions, creating a workforce proficient at using tools but less adept at critical thinking and problem-solving without them. This long-term impact presents a significant challenge, potentially hindering innovation and adaptability in the face of unforeseen circumstances.

The pursuit of streamlined efficiency, increasingly enabled by artificial intelligence, risks a subtle but profound shift in how professional work is valued – a process termed ‘Identity Commoditization’. This doesn’t signify a loss of jobs in the traditional sense, but rather the diminishing of unique skill sets and accumulated knowledge that define a professional’s identity. As AI systems take over routine tasks and even complex decision-making processes, the human element can be reduced to a readily interchangeable component, valued solely for the speed or cost-effectiveness of its output. Consequently, the deep expertise, critical thinking, and nuanced judgment that once distinguished a professional become less critical – and therefore, less valued – leading to a workforce where individuals are assessed not by what they know, but by how quickly they can execute instructions, ultimately eroding the very foundation of professional identity and long-term capability.

Building Sociotechnical Resilience Through Design

The proposed Dignified Human-AI Interaction Framework addresses potential negative consequences arising from increasing AI integration by focusing on the development of ‘Sociotechnical Immunity’. This framework posits that proactive design considerations, rather than reactive mitigation, are necessary to ensure beneficial human-AI collaboration. Sociotechnical Immunity is achieved by strengthening the interplay between human capabilities, social structures, and technological systems, effectively buffering against unintended consequences such as skill degradation, over-reliance on automation, and erosion of human oversight. The framework prioritizes maintaining essential human agency and control within systems increasingly mediated by artificial intelligence, acknowledging that complete automation is not always desirable or beneficial.

The implementation of ‘Do-Not-Automate Lists’ is central to maintaining crucial human capabilities within increasingly automated systems. These lists function as a proactive identification of tasks – typically those requiring complex judgment, ethical considerations, or unforeseen circumstance adaptation – that should remain under direct human control. By deliberately excluding these tasks from automation, organizations can preserve the cognitive and practical skills of their workforce, preventing skill degradation and ensuring a sustained capacity for effective response to novel situations. The creation of these lists requires careful analysis of workflows to differentiate between automatable processes and those where human oversight is non-negotiable for safety, accountability, or optimal performance.

Social Transparency, as a component of sociotechnical immunity, necessitates the clear communication of an AI system’s capabilities and, crucially, its limitations to end-users. This involves displaying information regarding the data used for training, the system’s known failure modes, and the confidence levels associated with its outputs. Implementing social transparency features allows users to accurately assess the reliability of AI-driven recommendations or decisions, facilitating informed oversight and preventing over-reliance on potentially flawed systems. Specifically, visibility into these constraints enables users to identify situations where human intervention is necessary, thereby maintaining critical skills and preventing deskilling, and allows for appropriate validation of AI outputs.

Radiation Oncology: A Case Study in AI-Augmented Expertise

Radiation oncology was selected as a focused case study to evaluate the influence of AI-assisted treatment planning systems on the skills of clinicians. This field presents a complex workflow requiring precise anatomical definition, dose calculation, and plan optimization, making it suitable for assessing the impact of AI tools on human expertise. The selection allows for a detailed examination of how reliance on AI affects a clinician’s ability to independently perform these tasks, identify potential errors, and adapt to novel clinical scenarios. The longitudinal nature of the study facilitates tracking changes in skill sets over time as AI integration evolves within the radiation oncology workflow.

A longitudinal study is currently being conducted to assess the subtle, long-term impacts of AI-assisted treatment planning on the expertise of radiation oncology professionals. This research focuses on identifying ‘asymptomatic effects’ – changes in skill and judgment that are not immediately apparent in clinical outcomes but may represent a gradual erosion of core competencies. The study employs repeated assessments of treatment plans generated both with and without AI assistance, alongside detailed analysis of clinician decision-making processes, to quantify potential shifts in expertise over time. Data collection includes metrics related to plan quality, treatment time, and the frequency of specific clinical reasoning patterns, allowing for a statistically rigorous evaluation of the effects of AI integration on specialist skills.

Human oversight, implemented through ‘AI-Off Reviews’, is a critical component of responsible AI integration in radiation oncology. These reviews involve clinicians independently evaluating treatment plans generated by AI systems without access to the AI’s recommendations or intermediate calculations. This process allows for the identification of discrepancies between human-generated plans and those produced by AI, serving as an indicator of potential skill degradation in clinicians who increasingly rely on AI assistance. Regular AI-Off Reviews establish a benchmark for maintaining clinical proficiency and, crucially, provide a safety net for detecting and correcting errors that might otherwise go unnoticed, directly contributing to patient safety and treatment efficacy.

The Imperative of Critical Technical Practice

The development of artificial intelligence demands more than simply technical proficiency; it requires a practice deeply informed by social awareness. This research highlights ‘Critical Technical Practice’ as an essential approach, one that actively integrates a robust understanding of the potential societal impacts of AI systems alongside core engineering skills. It posits that truly effective AI development necessitates anticipating and addressing the ethical, economic, and political consequences of these technologies, moving beyond purely functional considerations. By combining technical expertise with critical reflection, practitioners can proactively shape AI systems that align with human values and promote equitable outcomes, rather than inadvertently reinforcing existing biases or creating new forms of disadvantage. This interwoven skillset is crucial for building AI that genuinely serves humanity and fosters a sustainable future.

The development of a Dignified Human-AI Interaction Framework centers on proactively designing artificial intelligence systems that augment, rather than diminish, human capabilities. This framework moves beyond simple task automation, instead prioritizing the preservation of crucial skills and expertise within the human workforce. Researchers argue that successful AI integration requires a deliberate focus on collaborative partnerships, where AI handles repetitive or data-intensive aspects of work, freeing humans to focus on complex problem-solving, critical thinking, and creative endeavors. The ultimate goal is to build systems that not only improve efficiency but also empower individuals, fostering a sense of agency and maintaining the value of uniquely human contributions in an increasingly automated world.

The long-term viability of human expertise in an increasingly automated world hinges on a proactive approach to system design, one that prioritizes skill preservation and builds what researchers term ‘sociotechnical immunity’. This isn’t simply about resisting automation, but rather about strategically integrating artificial intelligence in ways that complement and enhance existing human capabilities. As demonstrated through qualitative frameworks and empirical evidence, systems built with this principle in mind exhibit a resilience to disruption; they don’t merely function, but actively safeguard the skills and knowledge of the people who interact with them. This focus on sociotechnical immunity suggests that thoughtfully designed AI can move beyond efficiency gains to offer a form of technological robustness, ensuring that valuable human expertise isn’t eroded by the very tools intended to assist it.

The pursuit of seamless human-AI interaction, as detailed in the exploration of the AI-as-Amplifier Paradox, necessitates a holistic understanding of systemic consequences. This article rightly focuses on the subtle erosion of skills-asymptomatic effects that undermine long-term worker resilience. It echoes Edsger W. Dijkstra’s assertion that, “Program testing can be effectively conducted by an intelligent person, but it is rarely done.” Just as rigorous testing reveals hidden flaws in code, this research advocates for proactively identifying and mitigating the hidden harms of AI-driven automation. Building ‘sociotechnical immunity’ isn’t simply about fixing isolated problems; it requires a structural approach that recognizes how changes in one area of a workflow reverberate throughout the entire system, affecting both productivity and human expertise.

Beyond Amplification

The concept of ‘sociotechnical immunity’ presented here feels less like a solution and more like a necessary articulation of the problem. The field has fixated on the apparent gains of AI-driven automation, often framing the technology as a simple amplifier of human capability. However, this paper correctly identifies the inherent trade-offs – the quiet erosion of skill, the subtle shifts in workflow that diminish human agency. The challenge isn’t simply building ‘better’ AI, but designing systems that acknowledge and mitigate these asymptotic harms. Good architecture, after all, is invisible until it breaks-and a broken workforce is a far more costly failure than a flawed algorithm.

Future work must move beyond measuring productivity gains and instead focus on the long-term consequences of task delegation. The focus on ‘dignified interaction’ is a start, but dignity is not a feature to be bolted on; it emerges from a system’s inherent respect for the capabilities it utilizes – or allows to atrophy. The true cost of freedom, it seems, is not the initial investment in technology, but the ongoing maintenance of human expertise, lest we find ourselves wholly dependent on the very systems intended to liberate us.

Ultimately, the field needs a more robust understanding of the feedback loops at play. Automation isn’t a one-time optimization; it’s a continuous process of renegotiation between human and machine. The current emphasis on AI-as-amplifier paradoxically obscures the fact that the most scalable solutions are rarely the cleverest – simplicity, and a clear understanding of system-level consequences, remain the most reliable path forward.

Original article: https://arxiv.org/pdf/2601.21920.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-31 00:47