A guide for researchers on how to access and analyze workforce impact data through the OpenAI Signals portal.
What OpenAI Signals Offers Researchers
OpenAI Signals is a portal for exploring workforce impact data related to AI. For researchers, it is a structured entry point: instead of scraping scattered reports or relying on anecdotal case studies, you work from a shared dataset designed around how AI adoption intersects with jobs, tasks, and skills. The value is less about a single headline number and more about consistent fields you can compare across roles, industries, and time windows.
Treat the portal as a research instrument, not a finished paper. The data supports questions about exposure, displacement risk, task automation potential, and skill demand shifts—but only if you define your unit of analysis clearly. Decide up front whether you care about occupations, tasks within occupations, or worker-level outcomes, and keep that frame stable as you query and export.
Access and Prepare a Workable Dataset
Start by confirming eligibility and any access requirements the portal documents (account type, research affiliation, terms of use). Read the data dictionary before you download anything. Note definitions for job titles, industry codes, skill taxonomies, and any flags that mark estimated versus observed values. Ambiguous labels are a common source of false precision later.
When you export, keep a fixed pipeline: raw file, cleaned file, analysis-ready table. Log every filter (geography, sector, occupation family, date range). Prefer reproducible scripts over one-off spreadsheet edits so coauthors and reviewers can reconstruct your sample. If the portal offers multiple related tables, join them on documented keys only—do not invent crosswalks from name similarity alone.
Analysis Approaches That Stay Defensible
Useful analyses usually fall into a few patterns. Descriptive maps show which occupations or sectors appear most exposed to AI-related task change. Comparative slices contrast groups (e.g., cognitive vs. manual task bundles, or roles with high vs. low digital intensity). Longitudinal views track whether exposure or skill demand moves in the same direction after adoption-related events—without claiming causal proof unless your design supports it.
- Define exposure operationally (task share, skill overlap, or role-level index) and stick to that definition.
- Separate “AI can touch this task” from “this job disappears”; the first is more often what the data can support.
- Report uncertainty: missingness, coarse occupation codes, and self-reported adoption all bias results.
- Pair Signals data with independent sources (labor statistics, firm surveys, task inventories) for triangulation, not for cherry-picking confirmation.
Avoid overclaiming. Workforce impact is multi-causal: demand shocks, offshoring, regulation, and firm strategy move with AI. Frame findings as associations under stated assumptions, and spell out what would falsify your interpretation.
From Portal Output to a Credible Research Product
Document methods at the same granularity you would for any public dataset: sample construction, exclusion rules, variable transforms, and robustness checks (alternate occupation groupings, stricter missing-data rules, leave-one-industry-out). If you publish figures, label axes with the portal’s field names and note the download date so others can align versions later.
Share code and a data-access recipe (which tables, which filters, which joins), not only charts. That practice turns a one-off portal session into reusable evidence for policy briefs, academic work, or internal workforce planning—while staying honest about what AI workforce data can and cannot settle on its own.