Summary: While public concern is widespread that artificial intelligence will dramatically displace workers, researchers argue that widespread job loss is unlikely. Still, integrating machine learning into the workplace could reshape the economy on a scale comparable to the steam engine and electricity.
Source: Carnegie Mellon University
Machine learning systems—algorithms that improve with experience—are set to reshape the economy much like past general-purpose technologies such as steam power and electrification. These systems can exceed human performance on many tasks, yet they are unlikely to replace humans across all jobs.
Researchers Tom Mitchell of Carnegie Mellon University and Erik Brynjolfsson of MIT present this view in a Policy Forum commentary to be published in the Dec. 22 edition of the journal Science. Mitchell, who established the world’s first Machine Learning Department at Carnegie Mellon, and Brynjolfsson, director of the MIT Initiative on the Digital Economy, outline a practical rubric of 21 criteria to evaluate whether a specific task or job is well suited to machine learning (ML).
“Although the economic effects of ML are relatively limited today and we are not on the brink of an imminent ‘end of work,’ the long-term implications for the economy and the workforce are profound,” they write. According to the authors, the skills individuals develop and the investments businesses choose to make will determine which workers and companies prosper as ML becomes embedded in daily life.
Machine learning is a core component of artificial intelligence. Rapid advances have produced notable gains in facial recognition, natural language understanding, and computer vision. ML is already widely adopted for credit card fraud detection, recommendation engines, and financial analytics, with emerging applications such as medical diagnosis gaining traction.
Predicting ML’s impact on any particular occupation is challenging because ML typically automates or partially automates discrete tasks, while many jobs comprise a mix of tasks—only some of which are suitable for ML methods.
“We don’t yet know how all of this will unfold,” Mitchell said. For example, earlier research showed a machine learning program could detect certain skin cancers more accurately than dermatologists. That result does not imply dermatologists will be replaced; dermatologists perform many functions beyond lesion screening, including patient counseling, complex diagnosis, and treatment planning.
“Dermatologists may become more effective and have more time to spend with patients,” Mitchell said. “Roles that center on human-to-human interaction will likely grow in value because those activities are hard to automate.”
The authors note several features that make tasks amenable to ML. Large, labeled datasets enable ML systems to learn complex patterns—skin cancer detection benefited from hundreds of thousands of annotated images; fraud detection systems are trained on hundreds of millions of transaction records. Tasks that already exist in digital form, such as scheduling or transaction monitoring, are easier to automate. Jobs that do not require advanced dexterity, mobility, or physical manipulation also tend to be better suited to ML.
ML excels at rapid decision-making when decisions rely primarily on patterns in data. Conversely, ML struggles when decisions require long chains of reasoning, broad background knowledge, complex context, or commonsense judgment.
The authors also emphasize that ML models often act as black boxes, producing accurate predictions without clear, human-readable explanations for how a conclusion was reached. For domains where a clear rationale is required—such as clinical diagnosis or regulatory decisions—a human professional still provides critical value by explaining reasoning and considering the broader context.
Work is underway to develop “explainable” machine learning systems that can make their internal reasoning more transparent and interpretable to users and regulators.
Source: Byron Spice – Carnegie Mellon University
Publisher: Organized by NeuroscienceNews.com
Image Source: NeuroscienceNews.com image credited to Tommy Leonardi
Original Research: Abstract for “What can machine learning do? Workforce implications” by Erik Brynjolfsson and Tom Mitchell in Science. Published online December 21, 2017. doi:10.1126/science.aap8062
Carniegie Mellon University. “How Machine Learning May Change Jobs.” NeuroscienceNews. Published December 22, 2017. Accessed December 22, 2017.
Abstract
What can machine learning do? Workforce implications
Over recent decades, digital computers have transformed work across virtually every sector of the economy. Today we stand at the start of an even larger, faster transformation driven by advances in machine learning (ML). ML can accelerate automation and serve as a general-purpose technology that enables many new applications and capabilities, similar to the historical impact of steam power and electricity. Yet there is no consensus about which tasks ML systems manage well, making it difficult to predict specific workforce and economic effects.
Using a rubric that outlines what current ML systems can and cannot do, the authors argue that many jobs contain components that are “suitable for ML” (SML) while other tasks within the same jobs are not. As a result, the impact on employment will often be nuanced: partial automation and task reallocation rather than wholesale replacement. Although ML’s economic effects remain limited today, the potential for rapid, large-scale changes—sometimes within a decade—means the technology could be highly disruptive, producing both winners and losers across the economy.
These developments call for careful attention from policymakers, business leaders, technologists, and researchers to manage transitions, update workforce training, and ensure broad-based benefits.