The comparison concerned deliverables rather than complete jobs
OpenAI introduced GDPval in September 2025 to compare AI-generated work samples with deliverables from experienced professionals. GPT-5 was among the models evaluated. The company described blind comparisons by occupational experts, not a trial in which a model independently performed every responsibility of a real job.
The announcement explicitly limited the first evaluation to one-shot tasks. It did not capture the whole process of clarifying an assignment, gathering missing context or revising work after feedback. Those limits matter when interpreting a claim about AI matching professional output.
The public records expose the task structure
OpenAI's public dataset contains 220 tasks across 44 occupations. Its records separate the written prompt, supporting reference files, example deliverables and scoring rubrics. The visible examples include work requiring spreadsheets and other files rather than only a short text answer.
The dataset card also explains that private individuals' identifying references are fictitious and that some material reflects sensitive themes found in real occupations. Those notes describe the research collection, not permission to treat a fictional scenario as a real customer case or a factual claim about a named person.
A strong work sample is useful evidence with a boundary
Comparing deliverables can reveal where a model helps with a well-defined piece of work. It cannot alone establish whether the same model can identify the right problem, manage a client relationship or take responsibility for an outcome. The release and dataset support a more concrete discussion than an unspecified claim of human-level work: a task has inputs, an expected artifact and a grading process. Whether those conditions resemble a particular workplace still requires judgment. These remain the developer's reported evaluations, not results independently reproduced for this article.