Yes—according to a January 9, 2026 WIRED investigation, OpenAI and training-data company Handshake AI reportedly asked third-party contractors for authentic workplace tasks and the files produced to complete them. The reported goal was to evaluate AI agents against human performance on complex, realistic assignments. Contractors were also told to remove personal, confidential, proprietary and material nonpublic information.
That does not prove that OpenAI trained ChatGPT on confidential company files, that every contractor uploaded former-employer work, or that any particular submission was retained or used. It does show why using real workplace artifacts as AI benchmarks creates difficult ownership, privacy, employment and trade-secret questions.
What contractors were reportedly asked to provide
WIRED said records from OpenAI and Handshake AI described requests for “real, on-the-job work” rather than summaries of a worker’s duties. The examples involved assignments that could take hours or days and were meant to represent the full chain from a request to a finished professional deliverable.
The task request
This is the instruction from a manager, colleague or client—for example, preparing a market analysis, building a financial model or writing a technical proposal.
#1 Best Overall
The task deliverable
This is the concrete output produced in response, not merely an explanation of what the worker did. Reported examples included Word documents, PDFs, PowerPoint presentations, Excel spreadsheets, images and code repositories. The list should not be read as an official, exhaustive file-type policy.
The human baseline
A completed human assignment gives evaluators something against which an AI agent can be measured: Did it interpret an ambiguous request, use multiple files or tools, sustain work across steps and produce an artifact that met professional expectations?
Where authentic material was unavailable, contractors could reportedly create fabricated examples designed to resemble realistic work.
WIRED reported these details on January 9, 2026; OpenAI and Handshake AI declined to comment to the publication.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhy real workplace data is attractive for AI evaluation
Synthetic prompts are easier to control, but they often omit the ambiguity and context that make professional work difficult. Real assignments can contain incomplete instructions, conflicting stakeholder preferences, unusual formatting, institutional terminology, messy spreadsheets and decisions that depend on earlier conversations.
Authentic work samples can therefore test whether an agent can:
- Turn an underspecified request into a workable plan.
- Find and reconcile information across documents, spreadsheets, presentations or software tools.
- Produce a usable business artifact rather than plausible-sounding text.
- Maintain accuracy over a multistep assignment.
- Match the quality, formatting and risk standards expected by a human professional.
The distinction between evaluation and training matters. Benchmarking measures an AI system against a reference result; evaluation tests quality and reliability; fine-tuning or training uses examples to change a model; and data analysis may study workflows or task structure. The available reporting describes evaluation of next-generation AI agents, but does not establish that every uploaded file was used directly to train a model.
What safeguards were reportedly offered
The reported instructions told contributors to remove or anonymize personal information, proprietary or confidential data, material nonpublic information, internal strategy and unreleased product details.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #3
WIRED also described a reference to a ChatGPT tool called “Superstar Scrubbing,” which supposedly offered advice on removing confidential information. The report did not independently establish the tool’s capabilities, deployment or effectiveness. A redaction assistant is not legal authorization to disclose a file and cannot guarantee that context has been made safe.
Why redaction does not resolve the rights problem
Removing names and logos may reduce obvious exposure while leaving the underlying work recognizable or restricted.
Ownership and authorization
A person’s possession of a file does not necessarily give that person the right to upload it. Control may depend on a work-made-for-hire provision, an intellectual-property assignment, an employment or contractor agreement, a client contract or a professional-services arrangement. A former employer or client may own the work, even if the worker created it.
Continuing confidentiality duties
Non-disclosure and security obligations commonly survive the end of a job. Uploading a past deliverable can raise questions about contract breach, client confidentiality and trade-secret protection even when the contributor no longer works for the organization.
Hidden and indirect identifiers
Potentially sensitive material can remain in:
- Unique project facts, dates, timelines and internal terminology.
- Document metadata, revision history, comments and tracked changes.
- Spreadsheet formulas, hidden sheets, named ranges and embedded files.
- File paths, code dependencies and screenshots of internal systems.
- Distinctive writing, formatting or combinations of facts that permit re-identification.
Personal and regulated information
Work files may contain customer, employee, patient, student, financial or contact information. Depending on the jurisdiction, industry and handling arrangements, healthcare, financial-services, education, government, legal, defense and human-resources rules may impose additional requirements. The report does not establish that any particular law was violated.
Intellectual-property lawyer Evan Brown told WIRED that the approach could expose contractors and AI companies to confidentiality and trade-secret claims. That is a risk assessment, not a finding that unlawful conduct occurred.
What the reporting establishes—and what it does not
| Question | What is supported | What remains unknown |
|---|---|---|
| Was there a request? | WIRED reported records describing requests for authentic workplace tasks and deliverables from third-party contractors. | Whether every contractor received the same request or was required to provide past-employer work. |
| What was the stated purpose? | To evaluate AI agents against human performance on realistic, economically valuable work. | Whether particular files were later used for model training, fine-tuning or another purpose. |
| Were safeguards mentioned? | Yes. Contributors were reportedly told to remove sensitive information. | Whether redaction was performed correctly, who reviewed files, and how effective any tool was. |
| Did confidential corporate data reach OpenAI? | The process created a risk that it could be submitted. | The contents, retention, access and disposition of individual submissions. |
WIRED separately reported an anonymous account from someone involved in selling failed companies’ assets: an OpenAI representative allegedly inquired about company data, potentially including documents, emails and internal communications, if personal information could be removed. The source said they did not pursue the idea. This account is distinct from the contractor-document evidence and is not independently verified in the available reporting.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Practical guidance for contractors
The safest rule is simple: do not upload a real work file unless the rights-holder has given clear written authorization and you have confirmed that no restricted information remains.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Do not upload without permission
- Client deliverables, employer-owned presentations or internal financial models.
- Unreleased product, engineering, marketing or strategy material.
- Code written for an employer or client.
- Customer lists, legal memoranda, medical or financial records.
- Government, defense or other NDA-covered work.
- Files containing third-party personal data or embedded confidential content.
Use safer substitutes
- Create a fictional document or synthetic task from scratch.
- Use public-domain or openly licensed material.
- Build a generic template with no client-specific facts.
- Submit only a task description, without the original deliverable.
- Transform an example only after the rights-holder approves the method.
Before submitting any authorized sample
- Read current and former employment, contractor, NDA and client agreements.
- Get written confirmation that the specific work may be disclosed for AI evaluation.
- Remove metadata, comments, tracked changes, formulas, hidden sheets, embedded files and internal paths.
- Ask how the file will be stored, who can access it, how long it will be retained and how deletion works.
- Keep the written authorization and the exact submitted version.
Do not assume that a scrubbed file is anonymous, that an old NDA has expired, or that a redaction tool guarantees confidentiality.
Practical guidance for employers and clients
- State in employment and contractor agreements whether work samples may be retained, placed in portfolios or uploaded to AI-evaluation programs.
- Define ownership, confidentiality, deletion and downstream-use terms for deliverables.
- Use technical controls to limit exports of sensitive files and preserve audit logs.
- Include AI-upload rules in offboarding and contractor-exit procedures.
- Review vendors’ provenance records, access controls, retention periods and incident processes before approving benchmarking projects.
- Offer synthetic or deliberately fictional alternatives for portfolio and evaluation use.
- Train staff that removing names is not the same as obtaining permission.
The broader AI-data question
The report sits within a growing market for specialized human data and professional task evaluation. WIRED identified companies including Surge, Mercor and Scale AI as part of that wider contracting and data-labeling ecosystem; that does not establish their participation in this specific OpenAI project.
The central issue extends beyond one company: AI developers need realistic work to test agents credibly, but contractors can become unauthorized channels for former employers’ or clients’ confidential information. Real files maximize realism; synthetic files provide clearer rights, repeatability and control. A responsible program may need both, with documented provenance, permission, structured redaction, access limits and retention controls.
Bottom line
The January 9, 2026 WIRED report supports a careful conclusion: OpenAI and Handshake AI reportedly sought authentic workplace tasks and deliverables from contractors to evaluate AI agents, while instructing contributors to remove sensitive information. The story does not prove widespread misuse or that confidential files were incorporated into a model. It does show that “scrub the document” is not a substitute for ownership checks, written authorization, confidentiality analysis and disciplined data handling.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
