Ethan Mollick — One Useful Thing · Zeitgeist

About #

Associate Professor of Management and co-director of Generative AI Labs at Wharton (University of Pennsylvania), and author of the bestselling book Co-Intelligence. Known for the “jagged frontier” framing of AI capability — the uneven, unpredictable boundary between tasks AI handles well and tasks it doesn’t. Writes research-grounded analysis of AI’s practical impact on work, management, education, and productivity.

2026-05-26 — Choosing to Stay Human #

Newsletter · Read

Core argument: using AI as a reflexive default risks eroding essential cognitive development — the choice of when and how to use AI matters more than whether to use it at all.
Research: a Turkish high-school math study found plain-ChatGPT users did better on homework but worse on unaided tests (the tool bypassed the learning effort); a Taipei Python course found personalized AI-tutored problem sequences lifted final exam scores 0.15 SD — equivalent to 6-9 months of extra schooling.
A BCG study of 758 consultants found AI users vastly outperformed peers overall, but failed far more often on problems where AI gave a confident-but-wrong answer — authoritative tone suppresses verification.
Anthropic research on programmers: those who let AI complete tasks outright couldn’t explain their own work, while those who requested explanations or used partial AI assistance retained understanding.
Practical takeaway: use “learning modes” in Gemini/ChatGPT/Claude, actively question AI outputs rather than accepting them passively, and consciously decide which cognitive tasks stay human — those decisions set lasting habits.

2026-06-04 — Co-Existence and the End of Co-Intelligence #

Newsletter · Read

Core argument: AI has shifted from a collaborative “co-intelligence” tool to autonomous agents that are sometimes — not always — better than the human; his new book Co-Existence addresses working productively with AI that isn’t a consistent partner.
Findings: AI coding agents now write roughly 80% of code at Anthropic and have increased output 17x, with developers shipping 8x more; capability remains “jagged,” excelling in some areas and failing unpredictably in others.
Framework — the “Authenticity Contract”: maintain genuine human authorship for credibility, since AI remains a poor long-form storyteller; advocates transparent, non-manipulative engagement with AI systems, even when marketing to them.
Practical takeaway: deploy AI selectively (fact-checking, feedback, unsticking creative blocks) while retaining core intellectual work; recognize AI as an increasingly important gatekeeper between creators and audiences, and treat AI collaboration as ongoing negotiation rather than a solved problem.

2026-06-09 — What it feels like to work with Mythos #

Newsletter · Read

Core argument: Claude 5 Fable (“Mythos-class”) marks a shift from “steering” AI to “commissioning” it — users act as patrons who lose visibility into the model’s decisions but gain access to dramatically greater capability.
Findings: Fable outperforms all prior public models across diverse tasks (academic papers, interactive games, math-rendered mapping apps) and autonomously spins up sub-agents to research, verify, and test — e.g., pulling 2,200+ flight/rail schedules for an isochrone map, or running 9.5 hours straight on a data-calibration tool.
Tradeoffs: token consumption roughly doubles versus prior models; output quality exceeds spec but still needs expert review for errors; safety guardrails are oversensitive and default to weaker models; the shift favors well-capitalized entities over independent developers.
Practical takeaway: human agency is moving from active steering to outcome validation — “the details of the AI’s decision making are not shown to me.”

2026-06-30 — The twilight of the chatbots #

Newsletter · Read

Core argument: AI capability is improving “better than exponentially” in autonomous task execution, and workplace usage is shifting from chatbot-style collaboration toward agent-based delegation.
Findings: METR and the UK AI Security Institute track programmer-equivalent-hours per prompt and find accelerating gains; Epoch found Claude Opus running autonomously for 14 hours could produce 2-17 weeks’ worth of human engineering for $251 in token costs; Mollick’s own tests had Fable running unsupervised for 9 hours, equivalent to team-weeks of work.
Frontier vs. near-frontier: US closed models (Anthropic, OpenAI, Google) lead, while open-weight Chinese models (Kimi K3, DeepSeek, Qwen) lag 6-12 months but follow their own exponential improvement curve.
OpenAI internal data shows 25% of workers run four or more agents weekly, with non-coding roles adopting agents at rates similar to engineers; research on Claude Code users found domain expertise — not job title — predicts who extracts the most value per interaction.

2026-07-23 — An opinionated guide to which AI to use to do stuff #

Newsletter · Read

Core argument: a practical decision guide for matching AI tool/mode to task — free/basic models for low-stakes queries, premium models (Claude Opus/Fable, ChatGPT’s GPT-5.6 Sol) for high-stakes work needing accuracy.
Recommends “Work”/“Cowork” modes (ChatGPT Work, Claude Cowork), which give agents virtual-computer access to complete hours of work autonomously, and Codex/Claude Code for full local-machine access on complex multi-file projects.
Security note: agent permissions matter — Claude requires approval before sending emails or deleting files by default, contrasted favorably against ChatGPT’s more permissive defaults.
Demonstrates an agent autonomously preparing MBA seminar materials via email access and web research; also flags specialized tools — Gemini Notebook for research, Gemini Omni for video editing.