How did Anthropic measure AI's "theoretical capabilities" in the job market?

Anthropic traced its “theoretical capability” baseline to an August 2023 report titled “GPTs are GPTs.”

The 2023 study used O*NET task breakdowns plus human and GPT-4-assisted labeling to judge whether LLMs or future LLM-powered software could halve task time.

Under that methodology researchers estimated about 15% of tasks could be halved and about 2.3% of occupations had at least half their tasks exposed to LLMs, and the authors warned of labeling subjectivity.