OpenAI Contractors Asked to Upload Work from Previous Jobs to Evaluate AI Agents

OpenAI is asking contractors to upload documents from their previous employers to help evaluate AI agents, according to a Wired investigation. The request signals both how aggressively the company is pursuing agent capabilities and how the data demands for next-generation AI systems are pushing into ethically murky territory.

This isn't about training models on scraped internet data—it's about testing whether AI agents can actually do real work. And apparently, OpenAI believes the best way to benchmark that is with the kind of documents you'd find on an actual employee's laptop: internal memos, project files, the unglamorous artifacts of professional life.

Why OpenAI Needs Real Work Documents for AI Agents

The distinction matters. Training data teaches models to predict text. Evaluation data tests whether an agent can accomplish tasks—respond appropriately to a customer complaint, draft a contract amendment, navigate a complex spreadsheet. Synthetic or public datasets can only take you so far. Real workplace documents contain the messy complexity that separates demo-ready AI from actually useful AI.

OpenAI has been racing to ship agent products, from the rumored "Operator" browser agent to enterprise tools that can take actions on behalf of users. To know if these systems work, you need to test them against the workflows they'll actually encounter. Hence: contractors uploading files from their previous gigs.

The IP and Confidentiality Problem

Here's where it gets uncomfortable. Those documents don't belong to the contractors uploading them. They belong to the contractors' former employers—companies that almost certainly have confidentiality agreements, trade secrets, and no idea their internal files are being used to evaluate OpenAI's products.

Most employment agreements include clauses about protecting proprietary information even after you leave. Uploading work documents to a third party would violate the spirit, if not the letter, of those agreements at many companies. The contractors may be getting paid by OpenAI, but they're potentially breaching obligations to past employers who never consented to this arrangement.

It's also worth noting what this reveals about data scarcity. OpenAI reportedly has access to vast training datasets, but evaluation data for agents is a different beast. You can't just scale up web scraping. You need representative samples of actual work—and those samples live in private enterprise systems that OpenAI can't easily access. So they're going through the back door: asking individuals to bring files with them.

What This Tells Us About Agent Development

The request underscores a broader truth about where AI development is headed. Models are increasingly good at generating plausible text. The bottleneck is increasingly about doing things—taking actions, completing workflows, operating in context. And evaluating that capability requires data that mirrors real-world conditions.

OpenAI isn't alone in facing this challenge. Every company building AI agents needs to answer the same question: how do you test whether your system can actually perform a job without access to real job materials? The temptation to cut corners on data provenance will only grow as the competitive pressure intensifies.

For now, the practice raises a straightforward question for anyone who's ever worked a knowledge job: what happens when the confidential files you created become someone else's test set? The answer, apparently, is that no one asks your former employer first.