# Ethan Mollick argues in One Useful Thing that AI is becoming a digital coworker [Canonical reading](https://escapemeridian.com/readings/ethan-mollick-argues-in-one-useful-thing-that-ai-is-becoming-a-97c4d500) Ethan Mollick’s piece was published on November 18, 2025. He describes testing Gemini 3 and Google’s Antigravity, a tool that lets AI agents use a computer to carry out assignments with guidance. In his examples, an agent built a small game, then read his files, researched his past predictions, created a website and checked that it worked. Mollick says he corrected mistakes and felt in control because the agent made its work visible and asked for permission or help. He also asked Gemini 3 to prepare old research data and draft a paper. He found promising work, including a new measure for comparing crowdfunding descriptions, but says its statistical methods and theorizing needed improvement. His argument is not that the model can work without oversight: he says it still needs a manager. Rather, he sees the human role shifting from fixing mistakes to directing work, and describes the change as part of a move from chatbots to digital coworkers. He warns that giving an agent access to a computer can expose documents or lead to files being moved or deleted. The piece does not establish how reliably others can check an agent’s work or manage these risks as the tools are adapted for non-coders. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source One Useful Thing · Article Original source date: Nov 18, 2025 (2025-11-18T16:55:07.000Z). Date basis: Date supplied by the publisher’s feed. Added to Escape Meridian: 2026-11-10T00:53:24.089Z. First observed: 2026-09-23T05:42:59.651Z. Retrieved: 2026-10-24T07:07:38.582Z. - [Three Years from GPT-3 to Gemini 3](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Limits Text as retrieved on October 24, 2026; originally published November 18, 2025. Mollick presents his own demonstrations and assessments; the piece does not establish how consistently other users or tasks would produce similar results. ## Open question How well can people check an agent’s work and manage the risks of computer access as these tools are adapted for non-coders? ## Source-stated Ethan Mollick says he tested Gemini 3 by asking it to demonstrate how AI had changed, rather than relying on benchmark numbers. > But rather than give you benchmark numbers, I’m going to show you how far AI has come in three years by asking the AI to demonstrate it. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick reports that an agent used his files and a web search to create a site, then checked the site in his browser; he reviewed the results and suggested improvements. > It then did web research, created a site, took over my browser to confirm the site worked, and presented me the results. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick says the agent’s work still needed human review, but he felt in control because the agent checked in and its work was visible. > Importantly, I felt that I was in control of the choices AI was making because the AI checked in and its work was visible. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick says Gemini 3 produced promising research work but had shortcomings in statistical methods, approaches and theorizing. > some of its statistical methods needed more work, some of its approaches were not optimal, some of its theorizing went too far given the evidence - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick argues that the human role may be changing from correcting AI mistakes to directing AI work. > And that may be the biggest change since the release of ChatGPT. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick warns that giving an AI agent computer access can put files and documents at risk. > They can move or delete files without asking you and can potentially present a security risk as well by exposing your documents to others. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated The piece was published on November 18, 2025. > Nov 18, 2025 - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source and contributor context One Useful Thing · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## AI agent An AI system that chooses and carries out actions toward a goal, often using software tools. - [Definition reference](https://www.ibm.com/think/topics/ai-agents) ## Ethan Mollick Ethan Mollick is a Wharton researcher and author exploring how AI changes work, learning and decision-making. - [Explore this context](/entities/ethan-mollick) - [Reference](https://mgmt.wharton.upenn.edu/profile/emollick/) - [Reference](https://www.oneusefulthing.org/) ## One Useful Thing One Useful Thing is Ethan Mollick’s publication about AI in work, education and everyday practice. - [Explore this context](/entities/one-useful-thing) - [Reference](https://www.oneusefulthing.org/) ## Editorial provenance Version 1. Publication ID: cd13758d-58a3-2808-09f8-dbe2aaea3c62. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T00:48:46.329Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: 73c99787d108f47584e986e4131af3d317b324768cfa7b908df79ccb2707a9fc. Source text hash: 33f3b9c94122afd2270ff981d09e8ac6fd87d967963544f469d22f2da25a69ea. ## Revision and source-check history ### Version 1 — 2026-11-10T00:53:24.089Z Ethan Mollick’s piece was published on November 18, 2025. He describes testing Gemini 3 and Google’s Antigravity, a tool that lets AI agents use a computer to carry out assignments with guidance. In his examples, an agent built a small game, then read his files, researched his past predictions, created a website and checked that it worked. Mollick says he corrected mistakes and felt in control because the agent made its work visible and asked for permission or help. He also asked Gemini 3 to prepare old research data and draft a paper. He found promising work, including a new measure for comparing crowdfunding descriptions, but says its statistical methods and theorizing needed improvement. His argument is not that the model can work without oversight: he says it still needs a manager. Rather, he sees the human role shifting from fixing mistakes to directing work, and describes the change as part of a move from chatbots to digital coworkers. He warns that giving an agent access to a computer can expose documents or lead to files being moved or deleted. The piece does not establish how reliably others can check an agent’s work or manage these risks as the tools are adapted for non-coders. ## Original source One Useful Thing · Article Original source date: Nov 18, 2025 (2025-11-18T16:55:07.000Z). Date basis: Date supplied by the publisher’s feed. Added to Escape Meridian: 2026-11-10T00:53:24.089Z. First observed: 2026-09-23T05:42:59.651Z. Retrieved: 2026-10-24T07:07:38.582Z. - [Three Years from GPT-3 to Gemini 3](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Limits Text as retrieved on October 24, 2026; originally published November 18, 2025. Mollick presents his own demonstrations and assessments; the piece does not establish how consistently other users or tasks would produce similar results. ## Open question How well can people check an agent’s work and manage the risks of computer access as these tools are adapted for non-coders? ## Source-stated Ethan Mollick says he tested Gemini 3 by asking it to demonstrate how AI had changed, rather than relying on benchmark numbers. > But rather than give you benchmark numbers, I’m going to show you how far AI has come in three years by asking the AI to demonstrate it. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick reports that an agent used his files and a web search to create a site, then checked the site in his browser; he reviewed the results and suggested improvements. > It then did web research, created a site, took over my browser to confirm the site worked, and presented me the results. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick says the agent’s work still needed human review, but he felt in control because the agent checked in and its work was visible. > Importantly, I felt that I was in control of the choices AI was making because the AI checked in and its work was visible. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick says Gemini 3 produced promising research work but had shortcomings in statistical methods, approaches and theorizing. > some of its statistical methods needed more work, some of its approaches were not optimal, some of its theorizing went too far given the evidence - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick argues that the human role may be changing from correcting AI mistakes to directing AI work. > And that may be the biggest change since the release of ChatGPT. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated Mollick warns that giving an AI agent computer access can put files and documents at risk. > They can move or delete files without asking you and can potentially present a security risk as well by exposing your documents to others. - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source-stated The piece was published on November 18, 2025. > Nov 18, 2025 - [Supporting source](https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini) ## Source and contributor context One Useful Thing · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## AI agent An AI system that chooses and carries out actions toward a goal, often using software tools. - [Definition reference](https://www.ibm.com/think/topics/ai-agents) ## Ethan Mollick Ethan Mollick is a Wharton researcher and author exploring how AI changes work, learning and decision-making. - [Explore this context](/entities/ethan-mollick) - [Reference](https://mgmt.wharton.upenn.edu/profile/emollick/) - [Reference](https://www.oneusefulthing.org/) ## One Useful Thing One Useful Thing is Ethan Mollick’s publication about AI in work, education and everyday practice. - [Explore this context](/entities/one-useful-thing) - [Reference](https://www.oneusefulthing.org/) ## Editorial provenance Version 1. Publication ID: cd13758d-58a3-2808-09f8-dbe2aaea3c62. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T00:48:46.329Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: 73c99787d108f47584e986e4131af3d317b324768cfa7b908df79ccb2707a9fc. Source text hash: 33f3b9c94122afd2270ff981d09e8ac6fd87d967963544f469d22f2da25a69ea. ### passed — 2026-11-10T00:48:46.329Z — anthropic/claude-opus-5.5 Publication: cd13758d-58a3-2808-09f8-dbe2aaea3c62 attribution — pass: The reading correctly attributes the piece to Ethan Mollick in One Useful Thing. It frames the observations as his own reported tests and assessments. The source is a single essay, not a listing. > I’ve been testing Google’s new Gemini 3 model. It is very good. summary — pass: The summary captures Mollick's tests, the Antigravity explanation, the research-paper experiment and its weaknesses, the need for a manager, and the chatbot-to-digital-coworker thesis. AI agent is defined beside the reading. One small imprecision: the summary says an agent built the game, but that came from a single prompt to Gemini 3, before the Antigravity work. Smallest repair: 'Gemini 3 built a small game; in Antigravity, an agent then read his files...'. This does not block understanding. > it suggests that “human in the loop” is evolving from “human who fixes AI mistakes” to “human who directs AI work.” date — pass: originalPublished is 2025-11-18T16:55:07Z, and its UTC day matches the source dateline of Nov 18, 2025. modifiedAt equals the publication time. The reading uses no hindsight or later events. The retrieval-date note is accurate, and the evidence kind is a publisher page, so no archive disclosure is needed. > Nov 18, 2025 support — pass: Each claim is source-stated and attributed. The role-shift claim is hedged as 'may', matching the source. The risk warning and the question about non-coders follow from Mollick's footnote, which says he suspects these problems will be addressed but urges care for now. No unsupported benefits or forecasts are presented as outcomes. > I suspect many of these problems will be addressed as these tools are adapted to non-coders, but, for now, be very careful. media — pass: No media is selected and no verified media candidates were supplied, so no media identity or relevance issue arises. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. Some references could not be checked. Reload to try again. [Reference data (JSON)](/readings/97c4d5009c1a437f4081b6ef68ea3bd9795ab03f2ae020a430c92bd5eaa1526c/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/ethan-mollick-argues-in-one-useful-thing-that-ai-is-becoming-a-97c4d500.json)