AI tools helped gather research and write the first draft of this article. I edited it, and every quote and number was checked against the original pages (One Useful Thing posts, the Harvard Business School working paper page, Wharton, Vox). The cover illustration is AI-generated. Details at the end.
Ethan Mollick is an associate professor of management at the Wharton School who writes about AI on a newsletter called One Useful Thing [1][2]. His posts are aimed at people who use AI at work and in classrooms, and they are useful in a way that a lot of AI writing is not: after reading one, you usually know what was tested, what was found, how far to trust it, and what you could try tomorrow.
That is not an accident of personality. It is a structure, and a structure can be studied. For a style study at Dapp.Expert, we analysed 24 of his posts from 2022 to 2026, about 43,600 words of his own text [S]. This article describes the pattern we found, using one of his best-known posts as the main example, and then asks a narrower question: which parts of the method can someone who writes about AI, finance or blockchain borrow honestly, and which parts only work if you actually did the experiment?
This is the eighth part of a series on voice and trust in editorial work. Earlier parts covered showing your evidence, disclosing how AI tools were used, correcting mistakes, attributing research, the difference between confidence and evidence, keeping an authorial voice when you use AI, and sentence rhythm. This one looks at how a single writer puts those pieces together.
The shape of a typical post
Across the 24 posts, the same six moves came up again and again [S]:
- A hook: a new model, an observation or a plain question.
- The problem, set up through research or the author's own use of the tools.
- A demonstration: what he tried, what he typed, what came back.
- A named idea that sums up the pattern.
- Practical advice for the reader.
- An open ending about where things are going.
None of these moves is unusual on its own. In our reading, what makes the posts work is the order, and the fact that the limits of each claim usually arrive next to the claim itself, not in a footnote at the end.
Case study: a result, then the nuance
The clearest example is "Centaurs and Cyborgs on the Jagged Frontier" (16 September 2023), in which Mollick summarised a study he co-authored with researchers from Harvard, MIT, Warwick and Boston Consulting Group [3].
The hook is a question readers already have. The post opens by saying that many people ask whether AI is really a big deal for work, and answers in the second sentence: "We have a new paper that strongly suggests the answer is YES." [3]
The result comes immediately, with numbers. Consultants who used AI finished 12.2% more tasks, completed them 25.1% faster and produced results rated 40% higher in quality than those who did not [3]. The study's own abstract on the Harvard Business School site gives the same 12.2% and 25.1% figures for 758 knowledge workers (BCG consultants), randomly assigned to work without AI, with GPT-4, or with GPT-4 plus a short prompting overview [4].
Then, before any explanation, come the limits. Mollick names the co-authors, points readers to the paper "if you have questions about numbers or methods", and says that he had to simplify 58 pages of findings to fit a post, so any mistakes are his. He adds that the experiments were preregistered but the paper was a new working paper, "not yet peer-reviewed" [3]. (The paper has since been published in the peer-reviewed journal Organization Science, in 2026 [4].) A reader learns how much weight the numbers can carry before learning what they mean.
The named idea comes right after the limits, and the evidence then tests it. Straight after the caveats, he gives the uneven pattern, where AI is very good at some tasks and quietly bad at others that look similar, a short memorable name: "what we call the 'Jagged Frontier' of AI" [3]. He illustrates it at once with a simple example: AI writes a good sonnet but struggles with a poem of exactly 50 words. The name is introduced early, but it is never left without evidence: the next sections supply the data.
The bad news gets the same space as the good news. After the gains on tasks AI handles well, the post turns to one task chosen so that AI would give a convincing but wrong answer. Without AI, consultants got it right 84% of the time; with AI, only 60-70% of the time [3]. The paper's abstract puts it as 19% less likely to produce a correct solution [4]. The concepts in the post's title, two ways of dividing work between people and AI, arrive only after both sets of results.
The post ends with caveats, a practical step and an open edge. In the closing section he notes that people "can go on autopilot" with AI and that its outputs were somewhat uniform [3]. The advice is modest: use AI on real work tasks often enough to learn where it is strong and where it falls short [3]. The last line looks forward rather than summing up.
Notice what the structure protects against. A reader who stops after the first few paragraphs still leaves with the numbers and the main caveats. A reader who only remembers the name still remembers that it describes an uneven result, not a universal one.
Five habits that carry the method
The case study shows the order. Across the wider set of posts, five habits keep showing up [S].
1. The experiment is described as a protocol, not a boast. Mollick often tests a tool himself and reports what he did. In a 2025 review of GPT-5 he writes that he kept typing variations of "make it better" and watched the model add features he never requested [5]. The value is not that he tried it. It is that the reader can see the input, the output and the gap between them. In our counts he uses "I" about 16.8 times per 1,000 words, against 6.4 and 9.8 for the two comparison writers in the study [S]. In our reading, much of that first person is method: what he typed, what he saw. The study marked a self-experiment in about a third (34%) of the coded fragments [S].
2. Uncertainty is stated, and usually located. Writing about Google's Gemini in 2024, he lists four possible explanations for a result, picks one, and adds: "I think 2 is most likely, but I have no idea if that is true." [6] In another post he writes plainly: "We don't know what the future really holds." [7] The reader can tell which sentences are findings and which are his judgement, because the judgement is marked.
3. A frame comes with its own limit. When he compares working with AI to working with an intern, he notes that the analogy is "very loose" and that human interns "should obviously not be treated like AI" [8]. When he proposes a way to decide whether to wait for better AI before starting a project, he follows it with: "Do not take Wait Calculations too seriously." [7] A frame that admits where it breaks is easier to trust than one that claims to explain everything.
4. Advice is conditional. Instead of telling everyone what to do, he ties advice to a situation: "If you assign problem sets as homework, you should test them with the latest AI." [9] Readers who are not in that situation can skip it; readers who are know exactly what to do next.
5. Numbers are estimates when they are estimates. "I suspect 5-10 hours of actively using AI on work tasks is the minimum," he writes about learning to use the tools [8]. That is a personal estimate, and the word "suspect" says so. In our counts his posts carry about 8.5 numbers and 9.2 links per 1,000 words, and his numbers more often describe a study or experiment than a market statistic [S].
His sentences also tend to be shorter than those of the other two writers in our sample: about 20 words on average against 24 to 25, with a median of 18 [S]. Short sentences make a step-by-step demonstration easier to follow, but they are a consequence of the method, not the method itself.
Where the method is weakest
The same study that found these strengths also found a weak spot. Many of Mollick's demonstrations are single sessions with a chatbot. They show what happened once, and they are presented as evidence, but nobody else can repeat that exact conversation [S]. The habit of describing the protocol softens this problem: you can at least see what he did. It does not remove it. A memorable anecdote can still carry more weight in a reader's mind than it deserves.
The named ideas have a similar risk. A short label is easy to repeat without the evidence behind it. In the case study above the name was backed by data within a few paragraphs; in a weaker post, a name could arrive without much underneath it. This is a general risk of the technique, not a criticism of any particular post.
There is also a point about how he uses AI in his own writing. His newsletter's About page says he writes every post himself and asks AI for feedback only after he has a complete draft [2]. In a 2024 interview with Vox he described using it to get unstuck, for example asking for 30 versions of a transitional sentence he could not get right [10]. Being open about that is part of why the method reads as honest. It also means that copying his structure while hiding your own use of AI would miss the point.
Borrowing the structure, not the experience
Our style study sorted his techniques into three groups [S]. Some transfer to any writer. Some only work if the writer really did the experiment. And some belong to him and should not be copied at all.
| Technique | Transferable? | Condition |
|---|---|---|
| Result first, limits right next to it | Yes | The limits must be the real ones from the source |
| Conditional advice ("if you..., then...") | Yes | Procedural steps only; no buy, sell or price advice |
| Stated, located uncertainty | Yes | Separate "known" from "our assessment" |
| A frame followed by its limit | Yes | Say where your own analogy breaks |
| One linking question per section | Yes | One per section, not a series |
| Self-experiment as evidence | Only with real experience | You actually ran the test and can show the log |
| Showing the prompt and the output | Only with real experience | Real screenshots or logs, labelled as an example |
| His named concepts, metaphors and closing lines | No | Use your own words for your own findings |
The middle rows matter most for anyone writing with AI help. The temptation is obvious: the self-experiment is the most persuasive part of the method, and an AI tool will happily write "I tested this for a week" for you. That sentence would be an invented experience, and earlier parts of this series explain why that breaks trust.
Here is what borrowing the transferable part looks like. The wallet and the numbers below are invented for illustration.
Without the structure: "This wallet is the safest option for beginners. Everyone should switch."
With the structure: "An independent security audit of [wallet], published in March, found no critical issues and two medium ones, both marked as fixed (report linked). That is one audit of one version; it says nothing about phishing or user mistakes. If you are choosing a wallet for small, everyday amounts, the audit is a point in its favour. If you hold larger sums, it is not a substitute for a hardware wallet."
The second version leads with a result, names its source, states its limit in the next sentence, gives conditional advice, and contains no invented experience. It would work equally well for a report on a new AI model, a DeFi protocol or a game economy.
What this article cannot tell you
Our style study is a working document, not peer-reviewed research. It covers 24 posts chosen by year and genre, not every post. The comparison is relative: we measured Mollick against two other writers in the same study, not against a reference corpus of English. Overall counts, such as first-person use and sentence length, are automatic corpus metrics. Features such as hedges, numbers, self-experiments and analogies were coded first by automated rules and then by two machine-run readings; the rules agreed only moderately with the readings, the two readings agreed closely, and no human editor double-coded the sample [S]. The posts themselves also contain prompts and AI responses, which affect some counts.
It also cannot tell you why his newsletter became popular. Structure is one plausible factor; his academic role, his book and the timing of his early posts on ChatGPT at the end of 2022 are others, and the study did not separate them [S].
How this article was made
| Task | What the AI did | What a person did | How it was checked |
|---|---|---|---|
| Research | Built the 24-post sample, counted features, drafted the style study | Chose the writer and the questions; set the rule that invented experience is excluded | Corpus list and metrics files kept with the study; limits stated above |
| Draft | Wrote the first draft from the study | Edited the text and the structure | Every quote and number compared with the live One Useful Thing posts, the HBS working paper page, Wharton and Vox on 15 November 2026 |
| Images | Generated the cover illustration; laid out the diagrams from the series template | Approved the illustration and the diagrams | Labelled as AI-generated in the disclosure line and alt text |
| Not done by AI | — | The decision to publish and responsibility for the text | — |
Conflicts of interest: none. Ethan Mollick was not involved in this article or the style study. Quotes are short excerpts for commentary; read the original posts for full context.
Responsible for this article: Dmitriy Lastovka, author. Corrections: contact form.
Sources
Labels: official page, newsletter post, research, interview, internal study.
- Wharton School, University of Pennsylvania, faculty profile: Ethan Mollick, Associate Professor of Management. Official page. https://mgmt.wharton.upenn.edu/profile/emollick/
- One Useful Thing, "About". Official page. https://www.oneusefulthing.org/about
- Mollick, E. "Centaurs and Cyborgs on the Jagged Frontier." One Useful Thing, 16 September 2023. Newsletter post. https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged
- Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper No. 24-013, September 2023; published in Organization Science, March–April 2026. Research (peer-reviewed). https://www.hbs.edu/faculty/Pages/item.aspx?num=64700
- Mollick, E. "GPT-5: It Just Does Stuff." One Useful Thing, 7 August 2025. Newsletter post. https://www.oneusefulthing.org/p/gpt-5-it-just-does-stuff
- Mollick, E. "Google's Gemini Advanced: Tasting Notes and Implications." One Useful Thing, 8 February 2024. Newsletter post. https://www.oneusefulthing.org/p/google-gemini-advanced-tasting-notes
- Mollick, E. "The Lazy Tyranny of the Wait Calculation." One Useful Thing, 16 January 2024. Newsletter post. https://www.oneusefulthing.org/p/the-lazy-tyranny-of-the-wait-calculation
- Mollick, E. "On-boarding your AI Intern." One Useful Thing, 20 May 2023. Newsletter post. https://www.oneusefulthing.org/p/on-boarding-your-ai-intern
- Mollick, E. "The Homework Apocalypse." One Useful Thing, 1 July 2023. Newsletter post. https://www.oneusefulthing.org/p/the-homework-apocalypse
- Vox, "What a practical approach to AI looks like, according to an expert" (interview with Ethan Mollick), 5 September 2024. Interview. https://www.vox.com/the-highlight/370246/what-a-practical-approach-to-ai-looks-like-according-to-an-expert
[S] Dapp.Expert style study: Ethan Mollick, 24 posts from One Useful Thing (2022–2026), 4 November 2026. Internal study (working document, not peer-reviewed).










