Particle.news
Download on the App Store

OpenAI Sought Real Workplace Deliverables From Contractors to Evaluate Next-Gen AI Models

The company is building human baselines for professional tasks to gauge progress toward more capable systems.

Overview

  • Records reviewed by Wired show OpenAI and Handshake AI asked contractors to upload actual files from past or current jobs, including examples like a real Bahamas yacht itinerary.
  • Instruction materials require removing or anonymizing personal and proprietary information and point contractors to a ChatGPT “Superstar Scrubbing” tool.
  • The work feeds an evaluation program started in September to compare model performance against human professionals across multiple industries.
  • An intellectual property lawyer warned the approach could trigger NDA and trade‑secret claims for both contractors and AI labs even if documents are scrubbed.
  • A source told Wired OpenAI also explored obtaining internal documents and communications from asset-sale firms handling data from defunct companies.