Documents: OpenAI is asking contractors to upload their work from current or previous jobs to evaluate its models, leaving it to them to scrub confidential info
To prepare AI agents for office work, the company is asking contractors to upload projects from past jobs …
Context & Ripple Effects
OpenAI’s effort to test models on real workplace tasks sits alongside an earlier report that Scale AI left customer training documents broadly accessible through shared links, underscoring how evaluation data can become a confidentiality exposure rather than merely a quality input. The company has also reportedly tightened internal protections for sensitive information, making the contractor-led collection process a notable governance boundary.
The request targets the gap between general-purpose model behavior and the varied documents, workflows, and judgments found in office work. It also arrives after FTC records demands concerning model risks and a security incident, giving data-handling choices added scrutiny.
First-order effects
- Contractors must decide what material can be safely submitted and bear the practical burden of removing employer or client confidential information before upload.
- OpenAI gains access to more realistic work artifacts for evaluating office-task performance, but the usefulness of that material depends on contractors’ redaction quality and permissions.
Second-order effects
- Organizations whose work may appear in submissions have reason to tighten confidentiality guidance and review contractual limits on reuse of work product.
- AI labs and evaluation vendors may face pressure to use clearer provenance controls, access restrictions, and auditable redaction processes, particularly after exposed training documents at Scale AI highlighted the operational risks of loose handling.
Third-order effects
- As workplace-agent evaluation moves from synthetic tests toward real work artifacts, data provenance and consent could become a competitive constraint on which firms can safely build credible benchmarks.
- If employers restrict reuse of work product, the sector may increasingly favor partnerships and controlled enterprise environments over contractor-supplied material; the available coverage does not establish which approach will prevail.
The trend: This is one instance of workplace-agent generalization becoming as much a data-governance problem as a model-capability problem.