Evidence from 2023 suggests that large language models had the technical potential to handle a substantial share of knowledge‑work tasks, but it does not pin down Jason’s specific claim well enough to call it clearly right or wrong.

Key points:

• OpenAI’s “GPTs are GPTs” paper (based on GPT‑4) estimated that, with access to an LLM, about 15% of all U.S. worker tasks could already be completed significantly faster at the same quality, and that when you include LLM‑powered tools, this share rises to roughly 47–56% of all tasks; around 19% of workers might see at least 50% of their tasks affected. (arxiv.org) These figures imply capabilities well in the range Jason is talking about, and often above 30% for some white‑collar roles.
• Goldman Sachs’ 2023 analysis found that generative AI could substitute up to about one‑fourth of current work tasks overall, with especially high exposure (25–50% of tasks) in many white‑collar occupations such as administrative and legal work. (gspublishing.com) That again puts a 30% task share within the plausible range of what LLMs could handle in principle.
• McKinsey’s 2023 report on the economic potential of generative AI concluded that current generative AI and related technologies had the potential to automate work activities that absorb 60–70% of employees’ time, with the biggest effects in knowledge‑intensive, language‑heavy activities (communication, documentation, supervision, etc.). (courses.cfte.education) This strongly supports the idea that a very large fraction of knowledge work tasks are technically automatable.
• Experimental evidence from MIT in mid‑2023 showed that ChatGPT substantially raised productivity for mid‑level professionals performing realistic writing tasks: time to complete tasks fell about 40%, while independent evaluators rated output quality about 18% higher, and judged it suitable for workplace use. (news.mit.edu) That supports the claim that LLM outputs can reach “practically usable” quality in real business contexts.

However:

• These studies typically measure potential or speed/quality gains on subsets of tasks, or aggregate exposure across all workers, not “what fraction of a typical white‑collar knowledge worker’s entire task mix can be done by LLMs at business‑ready quality by December 31, 2023.” Their task definitions, exposure thresholds, and populations differ, so none of them map cleanly to Jason’s “30% of knowledge work” figure.
• Some estimates focusing on LLMs alone (without full surrounding tooling) find lower immediate impact—on the order of 15% of tasks significantly sped up—only reaching 47–56% when assuming well‑designed software built on top of LLMs, which was still emerging in 2023 and unevenly deployed. (arxiv.org) That makes it unclear whether, as actually available and integrated by end‑2023, LLM‑based systems truly crossed a robust 30% threshold for the typical knowledge worker.

Given that credible 2023 research brackets the technically automatable share of tasks anywhere from about 15% (LLM alone) up to well over 50% (with supporting tools) depending on definitions and assumptions, Jason’s “30%” is plausible and directionally consistent with the literature. But because the available data don’t directly answer his specific, worker‑level claim, and the exact percentage is highly definition‑dependent, the prediction cannot be determined to be clearly correct or clearly incorrect.

Hence the verdict: ambiguous.