Ethan Mollick
American academic (Ethan R. Mollick), Associate Professor of Management at the Wharton School, University of Pennsylvania, where he co-directs the Generative AI Lab. The earlier research programme was on crowdfunding (the doctoral work and roughly a decade of subsequent papers on Kickstarter, equity crowdfunding, and entrepreneurial finance), then on educational technology and simulation-based teaching. The programme that matters for this topic is the post-2022 work on generative AI as a workplace and learning tool — running through the One Useful Thing Substack from 2022 onwards and consolidated in Co- Intelligence: Living and Working with AI (Portfolio, 2024).
Stake§
Mollick's stake combines the academic with the public- communicator role he developed substantially after the public release of ChatGPT in November 2022. The One Useful Thing Substack has been one of the most-read public-facing accounts of generative AI's effects on knowledge work, and the 2024 book is the consolidating statement. The empirical claims in the trade book are anchored in studies, including several Mollick and his collaborators ran themselves.
Co-Intelligence is the work that matters in this topic. The book's organising claim is that current large-language-model AI is a jagged frontier of capability — strong at some tasks that look hard (analysis, summary, code, certain kinds of writing), weak at some tasks that look easy (consistent arithmetic, dates, factual recall under pressure) — and that the productive relationship between human work and AI is closer to a collaboration with an unreliable colleague than to a tool or oracle.
The book's anchor empirical study is the Harvard Business School Boston Consulting Group experiment Mollick ran with several collaborators in 2023 (Dell'Acqua, McFowland, Mollick et al.). The study assigned 758 consultants tasks both inside and outside the AI's capability frontier, with random assignment to either no AI access, GPT-4 access, or GPT-4 access with training. For tasks inside the frontier, the AI-assisted groups completed 12% more tasks at 25% greater speed with quality 40% higher; for tasks outside the frontier — where the AI's outputs were plausible but wrong — AI-assisted consultants performed worse than the no-AI control, because they accepted the AI's suggestions when they should have rejected them. The result is the empirical core of the jagged frontier claim.
The book's implications for this topic are mostly about the feedback loop and deliberate practice. An AI tool that generates plausible responses at high frequency is a high-bandwidth feedback source in a way human teachers and peer readers cannot be — which makes it useful for the repetition with adjustment part of deliberate practice in domains where feedback was previously hard to come by. The same property makes the tool corrosive to the deliberate-practice condition when the practitioner is unable to evaluate the feedback's quality, which is the typical novice condition. The construct Mollick proposes for working with the tools — treating the AI as a co-worker whose outputs must be verified rather than as an authority — is in effect a procedural recommendation for keeping the feedback loop functional in the presence of a partner that produces both useful and confidently wrong output.
The construct's relationship to the transfer problem is the open empirical question of whether skill developed with AI assistance transfers to working without it. Mollick's position in the book is that the question is the wrong one — that the future of knowledge work substantially involves AI assistance, and the relevant skill is working with the tools rather than working independently of them. The position is consistent with the practical findings; the educational implications are still being worked out in the literature and in the policy of institutions trying to teach research, writing, and analysis to students with AI access.