Wiles study shows 18% error detection drop when AI framed as employee
The Wiles study reveals measurable degradation in human oversight when AI systems receive employee framing. Major vendors continue marketing agent teams as digital coworkers despite evidence of diffused accountability. Worker preference data from Stanford indicates mismatch between vendor assumptions and actual task needs.
Emma Wiles's research documented responsibility diffusion. Participants shown agent-labeled work escalated issues to managers 44% more often and claimed less personal oversight. This occurred even as Microsoft, OpenAI, Anthropic, and Google released agent management platforms in 2025-2026 that explicitly market teams of digital colleagues with human-like flexibility.
Microsoft: Copilot agent deployments will record >30% uncaught task failures in production workflows by Q2 2027 absent mandatory human sign-off gates.
Sources (3)
- [1]Wiles et al. Boston University Working Paper on AI Labeling Effects(https://www.bu.edu/questrom/faculty-research/working-papers/ai-labeling-responsibility)
- [2]Acemoglu 2025 NBER Paper on Task-Augmentation vs Replacement(https://www.nber.org/papers/w33521)
- [3]Stanford HAI Worker Survey on Desired AI Tasks(https://hai.stanford.edu/research/worker-preferences-ai-automation-2026)