OpenAI’s Ask From Contractors for Real Work Files to Test AI Raises Questions
OpenAI is asking third-party contractors to upload real assignments from their current or past jobs, as per a report in Wired. The goal is to test how well its next-generation AI models perform against humans in real work.
The project forms part of OpenAI’s push to build a “human baseline” for complex tasks. These benchmarks will help the company measure progress toward AGI, or systems that outperform people at most economically valuable work.
Contractors are told to submit two parts for every task. First, the original request, like an instruction from a manager. Second, the actual output they delivered. This could be a Word file, PDF, PowerPoint deck, spreadsheet, image, or even a code repository.
OpenAI wants work that reflects real, on-the-job experience. The instructions stress that tasks should be based on what workers have actually done.
How real-world work becomes AI tests
In internal guidance, OpenAI explains that long and complex assignments are the most valuable. These are projects that take hours or days. A sample task describes a luxury concierge manager creating a yacht trip plan for a family. The deliverable is a real itinerary prepared for a client.
Contractors can also create realistic mock examples. However, the company repeatedly emphasizes that genuine workplace tasks are preferred.
To reduce risk, OpenAI asks workers to remove personal details and confidential information. It even references an internal tool designed to help scrub sensitive data from files.
Yet, this is where concerns begin.
Legal risks and blurred boundaries
Intellectual property experts warn that even cleaned documents can carry hidden risks. Former employees often remain bound by nondisclosure agreements. Sharing internal formats, strategies, or workflows could expose trade secrets.
The process places heavy responsibility on contractors. They must decide what counts as confidential. If something slips through, AI labs could face legal trouble over misappropriated data.
This model depends on trust at scale. Thousands of workers are being asked to judge what they can safely share.
Why AI labs are doing this now
AI companies want models that can act like employees. That means handling emails, plans, reports, and business workflows. Synthetic data is no longer enough.
To bridge the gap, labs like OpenAI, Google, and Anthropic now pay skilled professionals to generate high-quality training material. This has created a fast-growing industry around AI data work.
Firms like Handshake AI, Surge, and Scale AI manage these contractor networks. The market is already worth billions.
OpenAI has also explored buying data from companies that shut down. These archives can include emails and internal documents. Some sellers have declined, citing concerns over privacy and incomplete scrubbing.
The message is clear. AI labs want real work, real context, and real complexity. That is how they plan to build systems that operate inside offices, not just chat windows.
