Can Inorganic Intelligence Do Useful Tasks?
For this research topic, GDPVal is used as a framework for useful tasks, even though GDP itself may not be the best measure of societal output.
What is GDPval?
GDPval measures real-world, economically valuable professional work, not academic puzzles.
Key properties:
- Measures knowledge-work productivity, not raw correctness
- Compares AI outputs vs experienced human professionals
- Uses blind expert judging (win / tie / loss)
- Focuses on deliverables, not multiple-choice answers
Scale:
- 44 occupations
- 9 economic sectors
- ~1,320 total tasks (~30 per occupation)
- Authored by professionals with ~14+ years of experience
Occupations Covered (Examples)
GDPval spans representative roles across the economy, including:
Professional & Technical
- Software developer
- Mechanical engineer
- Industrial engineer
- Project management specialist
- Computer & information systems manager
Finance & Legal
- Accountant & auditor
- Financial analyst
- Financial manager
- Personal financial advisor
- Lawyer
Healthcare
- Registered nurse
- Clinical nurse
- Healthcare services manager
- Medical secretary / admin
Sales, Marketing & Media
- Sales manager
- Real estate agent / broker
- Marketing & sales representatives
- Journalist / editor
- Producer / director
Operations & Public Services
- Compliance officer
- Administrative services manager
- Manufacturing operations supervisor
- Social worker
(44 total roles across 9 sectors)
What GDPval Tasks Look Like
**GDPval tasks are real work outputs, not trivia.
Typical deliverables include:
- Financial models and forecasts
- Accounting spreadsheets and variance analyses
- Legal briefs and contract risk summaries
- Engineering design documents
- Sales decks and pitch presentations
- Project plans and operational schedules
- Customer support workflows
- Editorial articles and media summaries
- Healthcare care plans
Tasks often include reference files (spreadsheets, datasets, templates) and require multi-step reasoning.
–
The Public “Gold Subset”
OpenAI has released a 220-task “gold subset” for public inspection and research.
Purpose:
- Transparency
- Reproducible evaluation
- Representative sampling of the full benchmark
These tasks:
- Come from the same 44 occupations
- Include full prompts and reference materials
- Are used in OpenAI’s Evals framework
Finance & Accounting Example
Auditor Sample Testing & Variance Analysis
Scenario: You are an auditor reviewing an Anti-Financial Crime risk dataset.
Required deliverables:
- Select a statistically valid sample (90% confidence)
- Perform Q2 vs Q3 variance analysis
- Add flags and calculations in new spreadsheet tabs
Skills tested:
- Statistics
- Spreadsheet manipulation
- Professional judgment
- Clear documentation
Finance Example
Profit & Loss Report for a Music Tour
Scenario: You are Finance Lead for a touring production company.
Required deliverables:
- Consolidate income, costs, and tax data
- Build a multi-sheet P&L workbook
- Write an executive-level financial summary
Skills tested:
- Financial modeling
- Data synthesis
- Business communication
How Outputs Are Judged
GDPval does not score answers as “correct” or “incorrect”.
Instead:
- Expert judges blindly compare AI output vs human output
- Judged on:
- Accuracy
- Completeness
- Professional quality
- Usefulness
- Results are reported as win / tie rates.
A reported score like ~70% GDPval means:
- The model’s output was judged as good as or better than a human professional in ~70% of tasks.
Why GDPval Matters
GDPval attempts to capture something traditional benchmarks miss:
- Real economic value
- Professional judgment
- Output quality, not token accuracy
- Productivity without direct labor input
It’s designed to answer:
Can this model actually do economically useful work?
Summary
- GDPval evaluates real-world knowledge work
- Covers 44 occupations, 9 sectors
- Uses expert-authored tasks and expert judging
- Includes 1,320 total tasks, with 220 publicly released
- GPT-5.2’s ~70% score reflects professional-level output parity, not exam performance
Appendix
Here are actual example tasks from the 220-task gold subset of OpenAI’s GDPval benchmark — the portion that’s been open-sourced so researchers can inspect real prompts and reference files.
Finance & Accounting
Auditor Sample Testing & Variance Analysis
Scenario
You’re an auditor reviewing a spreadsheet of Anti-Financial Crime risk metrics.
Deliverables include:
- Choose a statistically justified sample at 90% confidence
- Perform a Q2 vs Q3 variance analysis
- Add sample flags and calculations in new spreadsheet tabs
Clear instructions and structured output expected.
Profit and Loss Report for a Music Tour
Scenario
You’re Finance Lead for a production company’s fall tour.
Deliverables include:
- Consolidate income / cost / tax data across multiple sheets
- Build a P&L Excel workbook with detailed breakdowns
- Summarise results for executive review
Requires both numerical analysis and professional report structure.
Professional Services / Consulting
Many tasks require multi-step deliverables such as slide decks or written outputs based on real data and business context (not all prompts are publicly visible).
Examples include:
- Create market competitor landscape presentations
- Formulate strategic recommendations based on provided datasets
These combine analysis + narrative + structured formatting.
Where to Explore the Full Gold Subset
OpenAI’s GDPval gold subset (220 tasks with prompts and reference files) is available via:
GDPval is designed to answer: Can this system actually do economically useful work at a professional level?