Ethan Mollick argues in One Useful Thing that AI’s value depends on how people use it

Ethan Mollick argues in One Useful Thing that AI’s value depends on how people use it

Ethan Mollick’s piece was published September 29, 2025. He argues that AI can now perform economically valuable tasks, including reproducing research from existing data, but that doing tasks is not the same as replacing whole jobs. In his account, Claude Sonnet 4.5 worked through replication files and results for several papers; Mollick says spot checks and a second AI model’s reproduction checked out, while some papers could not be accessed. He sees a possible way to scale checks on published research, while noting barriers including benchmarking for accuracy and fairness. Mollick also warns that agents can multiply work people do not need: he asked Claude to turn a memo into presentations and received 17 versions. He argues that the important distinction is whether people use their judgment to decide what is worth doing, rather than simply producing more. He relays an OpenAI paper’s suggested workflow: delegate a first pass, review it, try corrections if needed, and do the work oneself if that fails. The paper estimates this could make work faster and cheaper while retaining human control. Whether AI can reproduce research accurately and fairly at scale, and how organizations will choose to use agents, remain open questions.

Publication: fe9c690e-c71c-ed07-4a0e-690d05193565

attribution — pass: The reading correctly attributes the essay to Ethan Mollick in One Useful Thing. It keeps his arguments and reported tests framed as his claims, and attributes the suggested workflow to the OpenAI paper he relays.

Real AI Agents and Real Work The race between human-centered work and infinite PowerPoints Ethan Mollick

summary — uncertain: The reading is largely faithful, but 'agents' is central to it and never explained. The definitions list covers only Benchmark. Smallest repair: add the source's own plain description, AI that is given a task and can plan and use tools such as coding or web search to complete it, or add a term-library entry. The reading also omits Mollick's opening evidence: the OpenAI expert-task test, where human experts won only narrowly, and the METR task-length trend. That trims context for his threshold claim but does not change his position.

The dream of autonomous AI agents, which, when given a task, can plan and use tools (coding, web search) to accomplish it, seemed far away.

date — pass: The instant 2025-09-29T18:52:42Z falls on the UTC day September 29, 2025, which matches the source dateline. There is no hindsight framing, and the kind is publisher-page, so no archive disclosure is needed. The modification date equals the publication date.

Ethan Mollick Sep 29, 2025

support — pass: Each claim is supported by the source text: the threshold and task-versus-job argument, the Claude replication with spot checks and a GPT-5 Pro cross-check, results on other papers, file-size and data limits, barriers around accuracy and fairness benchmarking, the 17 PowerPoints, and the judgment conclusion. The OpenAI workflow is presented as the paper's estimate. The question follows from the barriers the source names.

I tried this on several other papers with similarly good results, though some were inaccessible due to file size limitations or issues with the replication data provided.

media — pass: No media was selected, and no verified candidates were supplied, so no media is the appropriate choice.