The words are metaphors, useful to engineers as shorthand, but they become something else when they reach the public.
Across the technology industry, a quiet but consequential habit has taken hold: the naming of machine processes after the deepest textures of human inner life. When Anthropic called a new computational feature for its Claude agents 'dreaming,' a WIRED columnist paused to ask what is lost when engineers borrow the language of the unconscious to describe statistical operations. The question is not merely semantic — it touches on how societies form beliefs about intelligence, capability, and the nature of mind itself.
- Anthropic's decision to name a new Claude agent feature 'dreaming' has reignited a long-simmering debate about whether AI companies are systematically misleading the public through the words they choose.
- The terminology of human cognition — dreaming, memory, attention, reasoning — carries neurological and experiential weight that no current AI system can honestly claim, yet the industry deploys these words as standard marketing.
- Consumers, investors, and policymakers are making consequential decisions about AI based on a vocabulary designed to flatter the technology rather than describe it accurately.
- A WIRED columnist is calling for technically precise naming conventions, arguing that honest language would shift public discourse from awe toward accountability.
- The industry has little internal incentive to change — anthropomorphic naming drives investment and attention — but the growing gap between perception and reality may eventually force a reckoning.
The technology industry has developed a persistent habit: when engineers build something new into an AI system, they reach for the language of human experience to name it. Anthropic's recent introduction of a 'dreaming' capability for its Claude agents — a specific algorithmic technique framed in the vocabulary of sleep neuroscience — prompted a WIRED columnist to ask the industry directly why it keeps doing this.
The complaint is not about the technology itself, which is functional and reportedly useful. The issue is the naming. When a company calls a computational process 'dreaming,' it invites a particular misunderstanding — one where a reasonable person might imagine something resembling what happens in human sleep, rather than the statistical operation actually taking place. The terminology obscures rather than clarifies.
This is part of a broader pattern. The AI industry has spent years labeling features after human cognitive processes: attention, memory, reasoning, intuition. Each word carries baggage, suggesting a parallel to human neurology that does not hold. When a language model is said to 'reason,' it is performing pattern-matching on training data — outcomes may resemble reasoning, but the process is fundamentally different.
The stakes are real. Consumers, policymakers, and investors make decisions based on what they believe these systems can do, and anthropomorphic language systematically tilts those beliefs toward overestimation. It also creates conceptual drift: once a feature is called 'dreaming,' every subsequent clarification sounds like a correction of something that was almost true — when it was misleading from the start.
More technically precise naming would change nothing about how the features function, but it would change the conversation — shifting questions from 'can machines dream?' toward 'what specific process has been implemented, and does it solve the claimed problem?' That is a harder, more technically demanding conversation. It is also a more honest one. The pressure to have it is unlikely to come from within an industry that has found anthropomorphic language to be effective marketing — but the gap between public perception and reality has a way of closing on its own terms.
The technology industry has developed a habit that is starting to wear on at least one observer: when engineers build something new into an AI system, they reach for the language of human experience to describe it. Anthropic, the company behind Claude, recently introduced what it calls a "dreaming" capability for its managed agents—a technical process that the company has chosen to frame using terminology borrowed from sleep neuroscience and the unconscious mind. The move prompted a columnist at WIRED to ask the industry a direct question: why keep doing this?
The complaint is not about the technology itself. Anthropic's engineers have developed something functional and, by most accounts, useful. The issue is the naming. When a company decides to call a computational process "dreaming," it invites a particular kind of misunderstanding. A person who hears that an AI system dreams might reasonably imagine something closer to what happens in human sleep—the brain processing experience, consolidating memory, generating imagery from the depths of the unconscious. That is not what is happening. What Anthropic has built is a specific algorithmic technique designed to help its agents work through problems in a particular way. The terminology obscures rather than clarifies.
This is not an isolated incident. The AI industry has spent years naming features after human cognitive processes: attention, memory, reasoning, intuition. Each name carries baggage. Each name suggests a parallel to human neurology that may not exist. When a large language model is said to have "attention," it does not attend to things the way a person does. When it is described as having "memory," it does not remember in any sense that would be recognizable to someone who has forgotten where they put their keys. The words are metaphors, useful to engineers as shorthand, but they become something else when they reach the public.
The stakes of this naming practice are not trivial. Consumers, policymakers, and investors make decisions based on what they believe these systems can do. If the language used to describe AI capabilities borrows too heavily from human experience, those decisions may rest on a misunderstanding. A person might assume that an AI system with "reasoning" capabilities reasons the way humans do—weighing evidence, considering alternatives, arriving at conclusions through deliberation. In reality, the system is performing statistical operations on patterns in training data. The outcomes may resemble reasoning. The process is fundamentally different.
Anthropomorphic naming also creates a kind of conceptual drift. Once a feature is called "dreaming," it becomes harder to talk about what it actually does without first untangling the metaphor. Engineers have to spend time clarifying that no, the system is not actually dreaming, it is performing X operation. The public hears the clarification as a correction, which suggests the original claim was close to true. It was not. The original claim was misleading from the start.
What would change if companies adopted more technically precise naming? The features would function identically. Claude's agents would work exactly as well. But the conversation around AI capabilities would shift. Instead of asking whether machines can dream, people would ask what specific computational process Anthropic has implemented and whether it solves the problem the company claims it solves. That is a harder conversation, in some ways. It requires more technical literacy. But it is also a more honest one.
The pressure to adopt clearer naming conventions is unlikely to come from within the industry. Anthropomorphic language is effective marketing. It makes AI systems sound more capable, more human-like, more worthy of investment and attention. But the cost of that effectiveness is a public understanding of AI that is systematically skewed toward overestimation of what these systems are and what they do. At some point, that gap between perception and reality becomes a problem—not just for consumers, but for the companies themselves, when the gap eventually closes and the public realizes what they were actually sold.
Citações Notáveis
A WIRED columnist argued that AI companies should stop naming technical features after human cognitive processes, as the practice creates misleading impressions about AI capabilities.— WIRED opinion piece
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter what we call these features? If the technology works, isn't the name just a label?
Because the name shapes what people think the technology is. Call something "dreaming" and people imagine consciousness, processing, something happening beneath the surface. Call it "iterative refinement of token sequences" and they understand it differently—or at least they know they don't understand it, which is closer to the truth.
But engineers use these terms as shorthand among themselves. They know what they mean.
They do. But the names leak out. They end up in press releases, in product descriptions, in conversations with investors and regulators. Once they're public, they stop being technical shorthand and become claims about what the system does.
So the problem is marketing, not the technology.
It's both. The technology is real and functional. But the way it's presented to the world creates a false equivalence between what AI does and what human minds do. That gap matters when people are making decisions about how to use these systems or whether to trust them.
What would better naming look like?
Precise, unglamorous, honest. "Iterative problem-solving through constrained sampling." "Pattern-matching across training data." It's less exciting. It doesn't sell as well. But it's what's actually happening.