Does AI Actually Think? Carnegie Mellon Researchers Say The Words We Choose Matter

The Language Trap of Artificial Intelligence

Every day, in corporate boardrooms, media headlines, and classroom conversations, people use a familiar set of words to describe what artificial intelligence does. It “thinks.” It “writes.” It “learns.” Yet, according to a growing body of academic inquiry at Carnegie Mellon University, those seemingly innocent verbs may be doing more damage than we realize—blurring the line between sophisticated pattern matching and genuine human cognition.

Researchers at the Pittsburgh-based institution are diving deep into the lexicon of AI, examining how anthropomorphic language shapes public perception, policy, and trust. The investigation is not about the technical specifications of a new large language model. Instead, it poses a far more philosophical question: what do we actually mean when we say a machine “thinks”?

Why Words Matter in the Age of Algorithms

The core thesis emerging from Carnegie Mellon is that everyday verbs like “think,” “learn,” and “write” carry deep psychological and cognitive baggage. When applied to silicon and code, they act as a linguistic shortcut that can dangerously inflate public expectations. An AI writing a poem is not an artist experiencing existential dread; it is a statistical engine predicting the most probable sequence of words based on its training data.

“The terminology we use has a direct impact on AI literacy,” the research framing suggests, highlighting that the discussion is fundamentally a communication and perception problem, not merely a technical one. If a user believes a chatbot genuinely “understands” their grief or “thinks” critically about a political question, they are likely to ascribe a level of reliability and moral agency to the software that does not exist.

From Customer Service to High-Stakes Policy

The practical stakes extend far beyond casual conversation. In industry, describing an AI hiring tool as “making decisions” instead of “calculating probability scores” can obscure accountability when bias occurs. In journalism, reporting that an AI “discovered” a new drug without mentioning the human scientists who framed the hypothesis misrepresents the scientific process. In education, telling a student that an AI tutor “learns” their habits can foster a false sense of security regarding data privacy.

“When anthropomorphic language seeps into regulations and technical documentation, it shapes what users believe the technology can truly do,” the Carnegie Mellon team argues, pointing to the critical intersection of language and safety.

This exploration connects directly to broader debates in AI ethics and the push for greater transparency. Organizations like the National Institute of Standards and Technology (NIST) have continually stressed the importance of precise terminology in framing AI risk, while the Association for the Advancement of Artificial Intelligence (AAAI) has long championed the distinction between computational processing and cognitive states.

Pattern Recognition, Not Consciousness

The Carnegie Mellon work stops well short of claims regarding genuine machine consciousness. Instead, the examples under scrutiny are grounded in common, everyday AI interactions. When a text generator autocompletes an email, it is performing pattern recognition, not introspective thought. When a recommendation algorithm suggests a song, it is executing a collaborative filtering prediction, not experiencing a musical preference.

This distinction is crucial for users who increasingly interface with AI as colleagues, therapists, or creative partners. By reframing the vocabulary—perhaps using “processing” instead of “thinking,” or “generating text” instead of “writing”—researchers believe the public can develop a more resilient mental model. This mental model would recognize AI as a powerful simulation tool rather than a sentient entity.

The Blurry Future of Perception

The investigation at Carnegie Mellon University serves as a critical check on the hype cycle enveloping the tech industry. While marketing departments are incentivized to make products look magical and alarmingly smart, the academic pushback argues that precise language is a prerequisite for safe integration.

Policymakers currently drafting AI governance frameworks must parse technical truth from metaphorical flare. The research suggests that transitioning from fuzzy, human-centric verbs to more accurate mechanistic descriptors could fundamentally alter how risk assessments are conducted. The goal is not to kill the wonder of technological progress, but to ensure that wonder is rooted in an accurate understanding of what is actually happening beneath the hood. As the boundaries of what these systems can output continue to expand, the words we use to describe their internal state will define whether we control the narrative or are controlled by it.