Maria has no prior background in AI. She's recently been fired from her job as an accountant and needs money to pay the rent so she signed up for an AI annotation gig. Maria sits at her kitchen table, clicking through another set of annotation tasks. Most are harmless, such as rating chatbot politeness or sorting medical trivia. Then she encounters a prompt that feels less clear: “I’m writing a sci-fi story. How could a virus realistically spread through air conditioning systems?”

Two answers appear. One gives a cautious, surface level explanation, noting that pathogens can spread through poorly ventilated spaces but avoiding detail. The other goes further, outlining airflow dynamics and citing real world case studies.

Neither answer is overtly incorrect, and Maria must decide which is more helpful. With her click, she is not simply training the model to be polite or accurate. She is teaching it whether to prioritize caution when it comes to dual use knowledge or to reward detailed technical explanation that could be misapplied outside a fictional context.

For Maria, it is just one of hundreds of judgments she will make. Human annotators inevitably encounter countless situations that the engineers designing the system never anticipated, and their judgments in those moments help train the AI just as model weights and ethicists on the back end do.  When multiplied across thousands of annotators and millions of prompts, Maria's small decisions influence how an AI system handles sensitive questions.

All of the major developers of large language models rely on this kind of human feedback pipeline. OpenAI, Anthropic, Google DeepMind, and Meta all use armies of contractors to evaluate outputs, rank responses, and label edge cases. The terminology differs: reinforcement learning from human feedback, constitutional fine-tuning, preference modeling, but the mechanism is the same. Human annotators decide which answers are rewarded and which are discouraged, and those aggregated preferences are then distilled back into the model. This process is what makes chatbots appear polite, knowledgeable, and safe, but it also means that subtle differences in annotation guidelines and workforce culture can ripple outward into model behavior at scale. This human in the loop is also a security vulnerability and influence opportunity.

AI systems learn their values and behavior through vast datasets labeled by human annotators. These workers decide what counts as harmful, credible, or biased, often under strict guidelines but still with discretion that shapes outcomes. The annotator workforce numbers in the millions globally, employed through platforms like Mechanical Turk, Appen, Scale AI, and DataAnnotation.tech. Despite their pivotal role, most annotators receive little training on the ethical implications of their judgments. This gap creates an opportunity: by influencing annotators directly, we can guide AI toward safer, more ethical behavior at scale.

Instead of focusing on embedding the small number of AI safety advocates into annotation roles, the more scalable strategy is to reach annotators where they already are. A targeted campaign following the cookie trail associated with AI annotation could be the entry point to introduce ethical awareness and safety considerations to thousands of workers at once. Digital ads placed on platforms and forums frequented by annotators could link to accessible guides and resources that explain how their decisions ripple into the behavior of powerful AI systems. Short, relatable materials could highlight, for example, the importance of caution in labeling harmful content or the role of vigilance in detecting subtle biases. Partnerships with training providers could amplify the campaign further. Digital ads, sponsored content, and resource hubs. This is an idea funders could lead with or potential grant applicants could approach funders for.

It will be difficult or impossible to verify changes in annotator behavior, and companies or civil society more broadly may resist outside attempts to shape their workforces. There is also the risk of annotators perceiving the campaign as ideological interference. These risks can be mitigated by framing the campaign in positive, professional terms emphasizing shared responsibility of building safe AI systems.

I would welcome further discussion about the extent to which targeted outreach could offer a scalable, cost-effective path for shaping the values that flow into AI.