OpenAI Wants Your Real Work Files: The Hidden Strategy for Next-Gen AI Model Evaluation

Your past work reports might be the secret sauce for the next AGI. OpenAI is reportedly asking third-party contractors to upload real assignments and deliverables from their current or previous workplaces to evaluate the performance of its next-generation AI models.

Establishing the Human Baseline for OpenAI AI Model Evaluation

According to internal records obtained by WIRED, this project is part of a broader effort to create a "human baseline" for complex tasks. OpenAI believes that measuring its models against high-level human professionals is a key indicator of progress toward achieving AGI—an AI that outperforms humans at most economically valuable tasks.

Contractors are instructed to turn their "long-term or complex work" into tasks for the AI. This includes uploading actual files—Word docs, PDFs, PowerPoints, and even code repositories. One example provided in a presentation featured a 2-page PDF yacht trip itinerary created for ultra-high-net-worth individuals, used as a benchmark for "experienced human deliverable."

The Legal Minefield of Trade Secrets

The practice has raised red flags among legal experts. While OpenAI advises contractors to use a tool called "Superstar Scrubbing" to delete proprietary information, intellectual property lawyers warn that this is a risky gamble. Contractors sharing documents from previous employers could be violating non-disclosure agreements (NDAs) or exposing trade secrets.

This isn't the first time OpenAI has sought sensitive data. Sources indicate the company once inquired about purchasing internal communications and emails from failed firms, though the deal didn't proceed due to concerns over data scrubbing reliability.