OpenAI Is Asking Contractors to Upload Work From Past Jobs to Evaluate the Performance of AI Agents
Summary
OpenAI has been asking third-party contractors to upload real examples of work they completed in past or current roles so the company can use those deliverables to evaluate and benchmark AI agents against human performance. Contractors are instructed to submit both the task request (what they were asked to do) and the actual deliverable (the file or artefact produced). OpenAI tells contractors to anonymise or remove personally identifiable information and proprietary material, and also allows fabricated examples to demonstrate typical responses.
The effort forms part of OpenAI’s push to measure model performance on real-world office tasks as it develops next-generation AI agents. The reporting highlights legal and privacy concerns: lawyers warn that even scrubbed documents could expose trade secrets or breach previous employers’ NDAs, and one source said OpenAI explored acquiring company data from firms winding down, raising further scrubbing doubts.
Key Points
- OpenAI asked contractors to upload concrete outputs (Word, PDF, PowerPoint, Excel, images, repos) from past jobs to create human baselines for evaluating AI agents.
- Instructions emphasise submitting both the task request and the deliverable, and repeatedly state examples should be work the person has actually completed.
- Contractors are told to remove personal information, proprietary/confidential data and material non-public information before uploading.
- Legal experts warn this approach risks trade-secret misappropriation claims and potential NDA violations if sensitive information slips through scrubbing.
- AI labs increasingly hire skilled contractors and specialist firms to produce higher-quality, real-world training and evaluation data—creating a lucrative sub-industry.
- OpenAI reportedly inquired about buying data from companies being wound down, but some sellers declined over concerns that anonymisation couldn’t be guaranteed.
Why should I read this?
Short version: if you handle work documents, NDAs, or hire AI services, this is worth your five minutes. OpenAI’s way of collecting real job outputs to teach and test AI agents bumps straight into privacy, legal and trade-secret landmines — and that affects employees, employers and any organisation thinking about using agent-style AI for office tasks.
Context and relevance
This story sits at the intersection of AI development practices and corporate risk. As labs push models to perform complex, long-running office work, they need high-quality, real-world examples to benchmark and train agents. That demand has driven AI companies to rely on specialised contractor networks and data firms to source realistic deliverables.
For readers in legal, compliance, HR or IT roles, the article signals a growing exposure: even carefully anonymised data can leak sensitive information, and outsourcing the decision about what counts as confidential puts significant trust in contractors. Regulators and organisations tracking data governance, trade-secret protection and GDPR implications should take note. It also reflects a broader industry trend toward using realistic, proprietary-like data to build evaluation datasets — which raises both product-quality and ethical questions.
Author style
Punchy. This is one of those reports that matters beyond headlines: it flags concrete operational and legal risks that organisations and individuals will face as AI agents move into real office workflows. Read the details if you work with sensitive documents or advise on data governance.
Source
Source: https://www.wired.com/story/openai-contractor-upload-real-work-documents-ai-agents/