Most diagrams of AI pipelines for drug discovery hype progress and blind spots media pipelines from-text-prompt-to production asset jump straight from "data" to "model" as if the gap were filled by math alone. In reality, a large part of that gap is filled by people working through queues of tasks: tagging images, rewriting sentences, rating outputs, flagging harm. Their decisions shape how models behave, what they refuse to do, and who gets hurt when they fail. The current wave of "foundation models" did not remove human labor. It reorganized it, offloaded it, and hid it. The layer nobody draws on the architecture diagram is a workforce. ## From raw data to "clean" training sets Training corpora sound abstract: web text, code, documents, conversations. Getting from there to something a model can ingest at scale is work ai how teams actually repartition tasks between humans and models. Someone has to: Clean and normalize data. Remove obvious junk, broken encodings, duplicates, corrupted files. Filter harmful or illegal content according to evolving policies. That means reviewing text, images, and videos that include abuse, exploitation, self-harm, hate, and harassment. Attach labels. Classifications, spans, bounding boxes, entities, sentiment, topics. Anything that lets the model map inputs to targets. Build small high-quality datasets used to steer, fine-tune, or evaluate the model. This work is done in three main settings. In-house teams, usually small, focused on critical or sensitive labels. Business process outsourcing firms, often in lower-income countries, providing large pools of annotators under long contracts. Crowdwork platforms, where tasks are posted as micro-jobs and picked up by individual workers under piece-rate pay. The last two carry most of the volume. They are far from the labs that publish papers and keynotes. For many workers, the job looks less like "teaching AI" and more like a stream of small screens with vague instructions, tight timers, and opaque rejections. ## Reinforcement learning from human feedback as industrial process Reinforcement learning from human feedback has been marketed as "alignment." At the operational level, it is a labeling and rating pipeline. The basic loop is simple: A model generates several candidate outputs for a prompt. Human raters compare them or rate each one on criteria like helpfulness, accuracy, safety policy why governments care about your gpu cluster loss functions. Those ratings turn into a reward signal used to train a reward model. The main model is then optimized to produce outputs the reward model scores highly. Behind this loop sits a workforce doing tasks like: Rank these four answers from best to worst. Read this output and flag any policy violations. Rewrite this answer to be more accurate, more concise, more neutral. Decide whether this refusal is appropriate or overly cautious. These decisions encode norms. "Helpful" versus "harmful" is not a neutral label; it is a choice. The combination of guidelines, task design, and rater demographics determines how those choices land inside the model. Reinforcement learning from human feedback is often narrated as a technical advance. It is also a large-scale labor regime: thousands of people paid to repeatedly decide what "good behavior" looks like for a machine. ## Content moderation by another name Many labeling and RLHF tasks are, in effect, content moderation. Workers are asked to: Identify hate speech, sexual content, and threats. Recognize references to self-harm, suicide, and eating disorders. Flag attempts to solicit illegal acts. Label graphic descriptions of violence and abuse as disallowed. What changes compared to classic social platform moderation is the framing. Moderators on a social network are told they are enforcing community standards. Annotators in an AI pipeline are told they are "training models" or "rating responses," even if the content is similar. The psychological impact is similar too: Repeated exposure to distressing material. Pressure to work fast under tight time limits. Limited context about what they are seeing or why. Sparse access to mental health support, and sometimes gag orders that prevent them from talking openly about the work. The invisible cost of "safer" models is that someone has to absorb the raw stream of what the model should refuse. That someone is rarely mentioned in the product announcement. ## Geographies, wages, and power The geography of algorithmic labor is not random. A significant share of annotation and moderation work is done in regions where wage expectations are lower and labor protections weaker. English-speaking workers in East Africa, South and Southeast Asia, Eastern Europe, and Latin America are common in vendor case studies. The economic logic is straightforward. Models owned by companies in high-income countries generate value for those markets. A cost-sensitive part of the pipeline is shifted to cheaper labor markets via outsourcing contracts and platforms. Power asymmetry shows up in small details: Workers often do not know which company or model they are labeling for. Task instructions can be vague or contradictory, but appeals cost time and risk pay. Rejections of work are sometimes opaque; pay can fluctuate without clear recourse. Workers have almost no influence over the policies they are asked to apply. The gap between the value created by models at the top of the stack and the pay at the bottom is large. That gap is not a technical fact; it is a design choice in how AI supply chains are built. ## Bias, culture ai boom neurips icml status games, and the shape of behavior Who labels and how they are instructed has direct effects on model behavior. If most raters share a cultural background, language talking to computers still hard register, or political environment, their judgments of "polite," "acceptable," "harmful," or "unacceptable" will reflect that. The result is models that generalize those norms back onto users who do not share them. If instructions are narrow or poorly localized, raters in one context may be applying rules crafted for another. That can mean: Overzealous blocking of certain topics in regions where they are legal and socially accepted. Under-recognition of harms that are more visible to marginalized groups. Flattening of language variety into a narrow standard perceived as "neutral." Bias in training data is frequently discussed as a property of corpora. It is equally a property of annotation regimes. Every labeled example is an opinion backed by pay. For teams building AI systems reliability engineering, this reality has two implications: Technical tweaks alone will not fix bias that enters through human judgment. If you never see or engage with the annotators' working conditions and instructions, you are flying blind on one of the main sources of your model's behavior. ## Evaluation as more labor Even after models are trained, human labor does not disappear. It shifts into evaluation and red-teaming. People are asked to: Probe models for jailbreaks and prompt-injection vulnerabilities. Stress-test safety policies with adversarial prompts. Rate responses for quality against internal benchmarks. Manually inspect logs for worrying patterns. This is skilled work. It requires understanding both the model's capabilities and the company's risk profile. Yet it is often handled as another queue of tasks, outsourced or pushed onto overstretched internal teams. The same patterns appear: fragmented responsibility, limited context, and a sense of chasing issues rather than shaping the system. Related perspectives appear in our analysis in Fine-Tuning, Adapters, and Instruction Tuning: A Practical Map of the Options. ## Alternatives and improvements The question is not whether to use human labor. It is how to structure it. There are several levers available. You can move from anonymous crowdwork to persistent teams, with better training, clearer feedback, and the ability to specialize. That turns annotation into a job, not a disposable task. You can demand, in contracts with vendors, minimum pay standards, mental health support for exposure to traumatic content, and transparent appeals processes. Those requirements then propagate down the chain. You can involve labelers in feedback loops: collect suggestions on task design, unclear policies, and recurring edge cases. That turns them from passive executors into contributors to quality. You can diversify rater pools where it matters, especially for safety and normative judgments, and be explicit about whose perspectives are underrepresented. You can document the labor layer in your model cards and system descriptions: where labels came from, how raters were selected, what instructions they followed, what constraints they worked under. None of this is glamorous. All of it is more honest than pretending "the model learned this from the data" without mentioning who turned that data into something legible. ## Why this belongs in governance discussions AI governance discourse often jumps straight to high-level principles: fairness, accountability, transparency. Algorithmic labor is where those principles become concrete. Fairness is not only about the model's outputs. It is also about whether the people doing the invisible work are paid fairly and protected from avoidable harm. Accountability is not only about corporate boards and CEOs. It is also about who gets to set labeling policies and who bears the consequences when those policies fail. Transparency is not only about explaining model decisions to end users. It is also about acknowledging that "human in the loop" means specific humans under specific conditions. If governance frameworks ignore the labor that feeds and constrains models, they will misdiagnose where problems arise and where remedies should apply. ## Seeing the stack clearly The language of automation invites a particular illusion: that intelligence has been abstracted away from people and embodied in machines. Look closely at modern AI systems and the picture is different. You see: People writing and revising policies. People cleaning, filtering, labeling, and rating data. People reviewing edge cases, incidents, and failures. People absorbing the emotional cost of the content models are told to reject. The more scale these systems reach, the more the underlying labor matters. For quality. For ethics. For legal exposure. For basic decency. Ignoring that layer does not make it vanish. It simply keeps it off slide decks and out of budgets until something breaks in public. A realistic governance conversation about AI starts by putting algorithmic labor back where it belongs: not as a footnote, but as one of the main components of the system.