Stop Naming AI Features After Human Processes

The latest announcement from Anthropic at its developer conference has reignited a growing frustration among AI researchers, ethicists, and technology commentators: the industry's persistent habit of naming artificial intelligence features after distinctly human cognitive processes. The company unveiled a new capability it calls "dreaming," designed to help AI agents organize and sort through stored "memories." While the terminology may sound intuitive to consumers, it represents a fundamental misrepresentation of how these systems actually function.

This naming convention has become increasingly prevalent across the AI industry, with major companies regularly adopting vocabulary that borrows heavily from human neurology and psychology. When companies talk about AI "thinking," "learning," "remembering," and now "dreaming," they're using language that carries deep associations with consciousness, understanding, and intentionality. However, these terms obscure the mathematical and computational realities underlying these technologies. The algorithmic processes happening within neural networks operate on fundamentally different principles than the biological processes they're being compared to.

The problem extends beyond mere semantics or academic nitpicking. This anthropomorphic language shapes how the public understands and relates to artificial intelligence, potentially leading to overestimation of capabilities and misplaced trust. When a feature is called "memory," people naturally assume it functions similarly to human memory—selective, creative, prone to certain types of errors and distortions. In reality, what AI systems store and retrieve is data transformed through mathematical operations, bearing little resemblance to the associative, emotional, and contextual nature of human recall.

Consider the specific case of Anthropic's new "dreaming" feature. The term evokes images of an AI system engaged in some sort of subconscious processing, perhaps consolidating information or working through problems while in a quiescent state, much like human brains are thought to do during REM sleep. The reality is considerably more technical: the system is performing computational processes to organize and cluster information in ways that make subsequent retrieval and analysis more efficient. It's an optimization algorithm, not the philosophical equivalent of what happens when humans sleep.

The terminology problem becomes particularly acute when considering how these descriptors influence regulatory discussions and public policy. Policymakers and legislators often lack deep technical expertise in AI development and machine learning. When they encounter terminology suggesting that AI systems possess memory, consciousness, or dreams, it can inadvertently elevate their perception of the technology's sophistication and autonomy. This misunderstanding can lead to inappropriate regulatory frameworks—either too permissive because the AI is seen as more understandable and controllable than it actually is, or too restrictive because it's seen as more dangerous and sentient than the technical reality warrants.

Major technology companies have employed this strategy for years, though perhaps not always with cynical intent. Sometimes it's a matter of convenience—the human-derived terms are already embedded in our language and understanding. Other times, there may be a genuine belief that making AI capabilities sound relatable helps the public grasp their potential. But good intentions don't make the practice any less problematic. When OpenAI talks about ChatGPT having "thoughts" or when Google describes its algorithms as "learning" from data, they're trading accuracy for accessibility in a way that ultimately undermines both.

Some argue that the alternative—using purely technical terminology—would make AI harder for general audiences to understand. But this objection misses the point. The goal shouldn't be to simplify through metaphor; it should be to explain accurately without being unnecessarily obscure. Saying an AI system "processes information in ways that identify patterns and generate statistically likely outputs" is more precise than saying it "dreams" or "thinks," and it's not significantly more difficult to understand. The issue is that the industry hasn't made a concerted effort to develop clear, accurate language that's also accessible.

There's also a troubling commercial dimension to this linguistic choice. Companies have a vested interest in making their AI systems sound impressive, intuitive, and almost human-like. Anthropomorphic language serves marketing purposes—it makes the technology seem more advanced, more capable, and more worthy of investment. A feature called "dreaming" has more appeal to venture capitalists and potential customers than a feature described as "data clustering optimization through iterative algorithmic refinement." The incentive structure encourages continued misrepresentation.

The scientific and academic communities have been relatively quiet on this issue, perhaps because the terminology is so pervasive that resisting it feels futile. Yet scholars studying artificial intelligence ethics and AI safety have raised concerns about how anthropomorphic language affects public perception and policy discussions. There's a responsibility that comes with introducing new technology to the world, and part of that responsibility is communicating honestly about what it does and does not do. The field of AI is young enough that better naming conventions could still become normalized if the industry committed to them.

Moving forward, the AI industry needs to develop and adopt clearer terminology that accurately describes functionality without relying on human cognitive analogies. This might require collaboration between technologists, linguists, ethicists, and communications experts to create vocabulary that's both precise and accessible. Professional organizations could establish guidelines for feature naming that discourage anthropomorphic language in official documentation and marketing materials. Educational initiatives could help train journalists, policymakers, and the general public to recognize and question these linguistic choices.

The stakes are genuinely high. As AI technology becomes increasingly integrated into critical systems—healthcare, finance, criminal justice, national security—public understanding of how these systems actually work matters enormously. Misunderstanding their capabilities and limitations could lead to inappropriate deployment, misplaced trust, or inadequate safeguards. The terminology we use to describe AI isn't merely academic; it shapes how society relates to and regulates one of the most consequential technologies of our time.

Until the industry acknowledges and corrects course on this issue, the gap between technical reality and public perception will continue to widen. Companies like Anthropic should reconsider their feature naming conventions, opting for descriptions that illuminate rather than obfuscate. The work of building trustworthy, safe, and well-understood artificial intelligence systems requires honesty—and that honesty must start with how we talk about what these systems actually do.