The accuracy of artificial intelligence depends on the quality of its human trainers’ consensus. But what if that consensus is achieved not through reasoned debate, but through bullying? Insiders reveal that in AI training, social dynamics can trump objectivity, meaning your AI might be learning from the most domineering person in the room, not the most correct one.
The process is supposed to be collaborative. When raters disagree, they meet to align their judgments. However, one worker described the reality of these meetings as a situation where “the more domineering of the two bullied the other into changing their answers.” This means that accuracy can be sacrificed for the sake of resolving a dispute quickly and avoiding conflict.
This problem is not just anecdotal; it is a recognized flaw in group-based annotation. Sociologists who study digital labor confirm that individuals with stronger personalities or more motivation can unduly sway a group’s decision, skewing the results. The AI, in turn, is trained on this skewed data, internalizing the biases and errors that result from a flawed human process.
This raises serious questions about the integrity of AI’s knowledge base. We assume the information has been vetted through an objective process, but the reality can be far messier. The AI’s confident answer might not be based on facts, but on the outcome of a workplace power struggle, a hidden vulnerability in the foundation of artificial intelligence.