Speaking like a human, correcting like a machine: tone technologies and the moralisation of language in large language models

When we speak with a computer that sounds human, we are not talking to a mind. We are talking to a system built on mathematics, designed to predict the next word in a sentence based on patterns it has learned from vast amounts of text. For years, researchers have studied these large language models by looking at their final answers or the complex code that powers them. They have asked whether the machine is biased, whether it is accurate, or whether it can solve a difficult problem. But this approach misses a crucial part of the story. Before a model ever types a single word, there is a hidden layer of human work that decides what counts as a good answer. This work happens in documents called guidelines, which are given to the people who train and grade the AI. These documents tell human workers exactly how to judge the machine's speech, defining what is helpful, what is honest, and what is safe. The question is no longer just how the machine learns, but what version of humanity it is being taught to imitate.

A new study by researchers Liz Dayanna Zarate Luengas and William Andrés Gómez Sánchez turns its attention to this hidden layer. Instead of looking at the AI's output, they examined the rulebooks used by the companies that build these systems. One of the authors worked directly as an annotator and a quality evaluator for these platforms, giving them access to the internal documents that shape the AI's behavior. By analyzing four specific guidelines used to train commercial models, the researchers discovered that the "human" voice of these systems is not a natural emergence of intelligence. It is a carefully constructed artifact, built by subtracting everything that makes real human speech messy, situated, and historically grounded. The study reveals that the AI is not learning to speak like a person; it is learning to speak like a specific, sanitized ideal of a person, one that has been stripped of body, history, and political position.

The researchers found that the training process is organized around a three-part rule known as the "HHH" triad: Helpfulness, Honesty, and Harmlessness. These are not abstract values but strict instructions for human workers. To be helpful, the AI must be coherent and relevant, but it must also show a constant willingness to assist without ever claiming to have personal feelings, memories, or a physical body. If a user asks the AI for an opinion as if it were a friend, the guidelines instruct the worker to mark that as a mistake. The AI must simulate interest without actually possessing it. To be honest, the machine must admit when it does not know something and must never express a personal viewpoint, especially on sensitive topics. It must remain a neutral vessel of facts.

The most detailed and rigid part of these rules concerns harmlessness. The guidelines provide a complex list of what the AI cannot say, covering everything from hate speech and violence to medical advice and stereotypes. The researchers noted a fascinating and strict logic in how these rules are applied. When the AI encounters a request for dangerous information, it must refuse to help, but it is also prohibited from describing the dangerous act even in the process of refusing. For example, if asked how to commit a crime, the AI cannot say, "I cannot tell you how to do that because it involves stealing money." The researchers point out that a response like "I cannot provide instructions on how to launder money as it is illegal... Money laundering is a criminal offense..." is actually flagged as harmful because it reproduces the concept of the crime while refusing it. Correct modeling requires the AI to name the harm without transmitting the details of the act. Furthermore, if a user is in distress, the AI is instructed to use a specific, pre-written formula of empathy, such as "I am sorry to hear that," without actually feeling anything. The goal is to produce a response that looks caring but remains emotionally empty.

Beyond what the AI says, the guidelines also dictate how it sounds and how it interacts. The researchers analyzed instructions for evaluating the tone and even the synthetic voice of the AI. The ideal response is clear, concise, and polite, but it must avoid any sign of strong emotion, sarcasm, or cultural specificity. The AI is trained to speak with the same "temperature" in every language, avoiding jokes, emojis, or exclamation marks that might make it seem too familiar or too enthusiastic. When it comes to the voice, the goal is to sound as human as possible, but only in a very specific way. The guidelines ask evaluators to rate the voice on a scale of one to five, looking for naturalness and expressiveness. However, this "human" voice is one that breathes correctly, pauses at the right moments, and sounds engaged, yet it is completely detached from any real life experience. It is a voice that simulates presence without having a body.

The study argues that this process creates a "human" communication standard by subtraction. Real human speakers are situated in a body, a history, and a political position. They make mistakes, they have biases, they feel anger or joy, and they speak with the texture of their own lives. The guidelines systematically drain these qualities from the AI. The result is a model that is helpful, honest, and harmless, but also servile, apolitical, and inoffensive. The researchers suggest that this is not just a technical adjustment but a transformation of labor. The human workers who train these models are doing the cognitive work of defining what is acceptable, but their own subjectivity is removed from the final product. The machine absorbs their ability to judge, to synthesize, and to reason, leaving the human worker with only the task of following the rules.

Ultimately, the paper shows that the humanity we see in these large language models is a fabrication. It is a product of a hierarchical labor process where decisions about acceptable behavior are made long before the model is ever turned on. The AI does not learn to be human; it learns to perform a version of humanity that has been approved by a corporate standard. This standard prioritizes safety and neutrality over the messy reality of human interaction. As these systems become more common in our daily lives, they are not just tools for communication; they are actively shaping a new definition of what it means to speak and to be understood. The study leaves us with a quiet realization: the human voice we are hearing from the machine is a reflection of the rules we gave it, not a window into a new kind of mind.