Study Examines History of 'Strategic Ambiguity' in AI Terminology

The terminology used to characterize artificial intelligence—including everyday verbs such as "thinking," "learning," and "writing"—has long relied on a pattern of "strategic ambiguity" that overlaps technical definitions with human cognitive traits, according to new academic research first reported by TechXplore. In a study published in the IEEE Annals of the History of Computing, Carnegie Mellon University historian Christopher Phillips and University of Pittsburgh researcher Alison Langmead analyze the evolution of computer science vocabulary since the 1950s. The authors argue that double-meaning terminology frequently leads the public to misinterpret machine functions while allowing technical creators to convey complex systems through familiar concepts.

Phillips, a professor and head of the Department of History in CMU’s Dietrich College of Humanities and Social Sciences, questioned why society continues to apply human-centered vocabulary to computational devices. He observed that terms like "smart," "read," "write," and "think" carry vastly different meanings when applied to digital algorithms compared to human beings. The paper asserts that strategic ambiguity enables specialized words to possess precise, technical meanings for computer scientists while simultaneously suggesting a far broader, anthropomorphic capability to lay audiences. Langmead emphasized that because computational systems are now deeply enmeshed in daily social structures, public discourse must transcend current tech industry hype cycles to clearly establish what machines can and cannot perform.

Rather than focusing solely on current debates over whether machines possess humanlike cognition, Phillips and Langmead examined historical conversations from the 1950s and 1960s, a period before artificial intelligence became a household term and long before personal computers were widely deployed. During that era, early computing pioneers debated how to describe machine capabilities. For example, computer scientist Norbert Wiener described computers as "learning" when they executed specific rule sets that led to more successful outcomes. While fellow researchers understood the rigorous mathematical definition Wiener intended, broader audiences interpreted the word through the lens of human experience and intellectual development.

The study’s authors stress that identifying this linguistic phenomenon does not diminish the technical achievements of modern generative AI systems. Phillips characterized modern large language models as remarkable engineering developments capable of generating outputs that people instantly recognize as coherent and meaningful. However, the study critiques modern benchmark tests used to evaluate AI models, including the Massive Multitask Language Understanding (MMLU) benchmark and the recently introduced "Humanity's Last Exam." Phillips and Langmead contend that these evaluations primarily measure classification accuracy—how accurately a model selects correct options on standardized tests—rather than demonstrating authentic human knowledge, reasoning, or understanding.

The paper further cautions that describing artificial intelligence systems as "thinking" or "writing" can lead society to view software as a human competitor rather than a functional tool. Phillips cited image generation as an example, noting that when an AI system renders a prompt like a dog riding a pony in a New York Mets parade, the result represents a significant technical accomplishment rather than human creativity. The researchers argue that reducing complex activities like learning, reading, and artistic creation to algorithmic outputs overlooks the personal relationships, emotional depth, and lived human experiences that fundamentally shape those endeavors.

One of the paper's key historical findings was the deeply interdisciplinary nature of early computer science discourse. During the mid-20th century, historians, literary scholars, engineers, psychologists, and computer scientists routinely collaborated to determine the societal boundaries and purpose of computing technology. Commenting on the study, Andreea Ritivoi, the William S. Dietrich Professor of English and associate dean of research in CMU’s Dietrich College, noted that language does not merely convey scientific ideas but actively alters human perception of them. Ritivoi called for renewed cross-disciplinary collaboration, referencing C.P. Snow’s classic work on the rift between science and the humanities.

Rob Kass, Maurice Falk University Professor of Statistics & Computational Neuroscience in CMU's Department of Statistics & Data Science and School of Computer Science Machine Learning Department, echoed the paper's emphasis on linguistic precision. Kass noted that while computational algorithms and statistical derivations require absolute exactness, everyday language relies on flexible words with context-dependent meanings. Kass stated that Phillips and Langmead convincingly demonstrated how ambiguous phrasing first emerged in early AI research and continues to be deployed for strategic advantage, ultimately impacting broader public understanding of the technology.