Dark Side of the AGI Moon (www.japantimes.co.jp)

AI systems that can draft sensitive documents — like autopsy reports — don’t arise from code alone but from large-scale, often invisible human labor: workers who sort and annotate thousands of crime-scene and corpse images so models learn terminology, structure and judgment. That precarious, low-paid work (often “a few dollars” per task) has prompted organizing from Kenya to Colombia, with workers describing severe psychological strain and no employer-provided mental-health support — “You have to spend your whole day looking at dead bodies and crime scenes,” one Kenyan annotator said. For the AI/ML community this exposes a fraught supply chain risk: the quality, bias and provenance of training labels for sensitive tasks are tied to exploitative labor practices and high emotional cost, which can degrade dataset integrity and create legal, ethical and reputational liabilities. Technical implications include increased likelihood of annotation errors, underreported edge cases, and downstream harm from models trained on traumatizing material. Solutions likely to gain attention are trauma-informed annotation workflows, fair pay and protections, clearer data provenance and auditing standards, and increased use of privacy-preserving or synthetic alternatives — all of which affect cost, model fidelity and deployment timelines for systems handling forensic or other sensitive content.