Written by: Rhonda Fletcher
At first glance, an AI assistant’s voice can feel inevitable — the neutral, politely informative cadence of a machine. But that voice is not a by-product of algorithms alone. It is the product of human choices: dataset curators, labelers, policy writers, corporate product managers, and a constellation of social norms they bring to bear.
In short, AI tone is culturally coded. This investigation traces how human labor and cultural frames shape the way AI speaks, why that matters for trust and harm, and where the tensions lie between effectiveness, ethics, and power.
Human Scaffolding in “Unsupervised” Learning
Large language models are frequently described as trained on vast, “unsupervised” corpora scraped from the web. Yet, as the InstructGPT team made plain, unsupervised pretraining is only the starting point.
To make models follow instructions and behave “helpfully,” researchers fine-tune them with human demonstrations and preference rankings, a technique known as Reinforcement Learning from Human Feedback (RLHF).
As Ouyang et al. wrote in the seminal InstructGPT paper: “fine-tuning with human feedback is a promising direction for aligning language models with human intent.” *(Ouyang, 2022)
That simple fact reframes the problem: the tone and apparent values of a deployed model mirror the aggregated judgments of the humans who judged model outputs during fine-tuning.
But who are those humans? They are an army of tens of thousands, if not more, spread across the globe. What is their language, region, education, political leanings, and even level of fatigue at the end of a labeling shift? All of these factors can subtly shift a response from “authoritative” to “hesitant,” from “wry” to “dry,” and from “inclusive” to “exclusionary.”
Labelers: Not Just Annotators but Culture-Makers
The role of human labelers is often minimized in public discourse, but ethnographic and experimental work shows that labeling is an interpretive act. In a WIRED feature recounting a project to curb online harassment, researchers assembled a small, diverse team of labelers who debated ambiguous cases until they reached consensus, ultimately improving the model’s performance on misogyny detection. The takeaway: Cultural context and explicit deliberation matter to how a model learns to classify and respond.
From a cultural coding perspective, labelers encode norms of politeness, safety, and acceptable phrasing. A labeler in one country might mark a sarcastic phrase as harmful; a labeler in another might flag it as harmless wit. Those micro-decisions accumulate into macro-patterns: the model learns which registers to favor when speaking to an English-language user, which metaphors to avoid, and which assumptions to make about authority, expertise, and deference.
Alignment: A Double-Edged Sword
RLHF promises alignment. Alignment is generally defined as “actions that ensure an AI is developed and deployed in a way that aligns with human values and goals. The consensus holds that alignments produce fewer toxic outputs and more accurate task completion. But it also concentrates normative power.
The InstructGPT results themselves are telling: a smaller, human-fine-tuned model could be preferred to a much larger unfine-tuned model, not because of superior intelligence but because it had been guided to match human expectations. “Outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3,” the NeurIPS paper notes. (Ouyang & Wu, 2022)
That preference, however, is not value-neutral; it reveals whose values the training process privileges..
Critics have warned that this concentration of cultural decisions poses risks. The 2021 ACM paper “On The Dangers of Stochastic Parrots” argued that scale and opacity can hide harm. LLM models can reproduce stereotypes and erode linguistic diversity, while suggesting that human oversight alone cannot fully immunize systems against these problems. So Emily Bender et al. asked, “What are the possible risks associated with this technology?” They were uniquely interested in ”who benefits” and “who is silenced.” (Bender, 2024)
Tone, Trust, and the Politics of Voice
Why does tone matter beyond aesthetics? Because voice mediates trust. Experiments on human-AI cooperation show that users respond differently to agents depending on perceived identity and style: anthropomorphic or authoritative language can increase compliance, while overly humanlike tones can trigger skepticism among users who expect transparency about machine status.
Tone, therefore, becomes a lever for influence. If the model’s cadence appears intimate or deferential, users may attribute intentions or expertise that the model does not possess. Recent human-LLM cooperation studies at The Hong Kong University of Science and Technology underscore how interaction dynamics — including tone and explanation style — change user trust and cooperation.
This is not merely theoretical. Studies comparing AI-generated and human-authored texts find measurable differences along dimensions like stance, deixis, and hedging; those linguistic features affect persuasiveness and perceived credibility.
Ultimately, when designers intentionally cultivate a “friendly expert” persona, they are making a social choice about how much authority an algorithm should wield in everyday life.
But who codes our culture?
Labor, Industry, Governance, and Opacity
Investigative attention should therefore turn to three interlinked sites of power: labor practices, corporate policy, and technical opacity. Labeling work is often outsourced, underpaid, and invisible; yet it is where key cultural judgments are forged. Corporate policy teams decide red lines and guardrails; product managers decide whether a model will “sound” conversational or formal. And the models themselves remain black boxes: internal vectors, strategic fine-tuning steps, and aggregated labeler decisions are rarely public, making an independent audit difficult.
In a recent podcast, ArthroPod, anthropologist Stefan Hornemann argued that this is precisely where cultural analysis can help. Rather than treating algorithms as neutral artifacts, researchers urge attention to the social processes that create them — what one forum calls “anthropology and algorithms.”
The Call to Action intends to render visible the rituals and negotiations that produce a model’s comportment.
Accountable Cultural Coding
If human choices shape AI tone, then accountability must begin with human-facing processes.
Three practical reforms deserve priority:
- Labeler transparency and diversity reporting. Companies should disclose the geographic, linguistic, and demographic makeup of teams that produce human feedback and publish summaries of labeling protocols. As the WIRED example shows, deliberate, diverse deliberation improves outcomes.
- Open auditing of tone and style. Auditors should be able to test models across registers and cultural norms, measuring whether a model privileges certain dialects, idioms, or deference patterns over others. Peer-reviewed benchmarks for “tone fairness” could mirror existing fairness metrics for accuracy.
- Participatory policy design. Communities likely to be affected by an application — marginalized language groups, cultural minorities, or sectors like healthcare — should have a seat in deciding the voice and guardrails of systems that will address them.
These steps are modest compared to the scale of contemporary systems, but they point to a principle: cultural coding should be democratic, not opaque.
AI tone is not background noise. It is an interface where labor, culture, and corporate strategy meet, with tangible effects on trust, persuasion, and harm. As we delegate more social tasks to machine voices — customer service, tutoring, legal triage — we should ask not only whether the answers are correct, but whose voice is answering, and why.
The human hands that shape those voices cannot be erased by claims that show a preference for the “unsupervised” scale. An AI’s cultural code is written in human ink, and accountability requires us to read it.