AI tools helped analyse the texts, gather the research and write the first draft of this article. I edited it and checked every quote against the original page. The cover illustration is AI-generated. Details at the end.
In crypto, criticism and promotion often look the same on the page. One post says a project is a scam; another says it is the future. Both are confident and both are opinions, unless the reader can see what each one rests on.
Molly White is one of the best-known critics of the crypto industry. She writes the newsletter Citation Needed and runs the incident tracker Web3 is Going Just Great. Her conclusions are often harsh. Yet in our study of three analytical writers, her texts were the most heavily sourced: the most numbers, the most links and the only footnotes. Her criticism is built so that her judgements point back to documents the reader can check.
This is the tenth part of a series on voice and trust in editorial work. Earlier parts covered showing evidence, disclosing AI use, correcting mistakes, attributing research, keeping confidence in proportion to evidence, keeping an authorial voice when you use AI, sentence rhythm, and how Ethan Mollick and Benedict Evans build their explanations. This one looks at the third writer in our study and her main tool: the link between a claim and its source.
What the numbers show
Our style study measured 78 public texts: 24 by Ethan Mollick, 28 by Benedict Evans and 26 by Molly White (2022–2026, about 64,000 words of her own writing). The same script counted the same features for all three [1].
| Per 1,000 words (average per text) | Mollick | Evans | White |
|---|---|---|---|
| Numbers | 8.5 | 9.6 | 17.6 |
| Links | 9.2 | 3.1 | 14.7 |
| Footnotes | 0 | 0 | 4.7 |
| Hedges ("might", "perhaps") | 6.1 | 9.7 | 5.7 |
| Boosters ("clearly", "definitely") | 7.5 | 9.2 | 4.0 |

Two things stand out. White uses about twice as many numbers as the other two, and she is the only one of the three who uses footnotes. And she has the fewest boosters by a wide margin, and slightly fewer hedges than Mollick. Her sentences seldom push a claim with "clearly". The weight of a statement comes from the number and the source attached to it, not from the modal words around it.
A second measure points the same way. The study also coded a sample of short fragments from her texts. In two independent machine readings, 47% and 49% of her fragments contained an explicit attribution to a source: a document, a court, a regulator, a named person [2].
How far these numbers go: three writers, compared with each other, not with general English. Counts are automatic, and the footnote count is zero for Mollick and Evans because their formats do not use footnotes, not because they never cite. The coding was done by rules and an AI agent, not by two human coders.
One passage, the whole chain
The clearest example is her report from the fourth day of the FTX trial in October 2023 [3]. The passage about the exchange's "insurance fund" moves in a fixed order.
First, the testimony, attributed. She notes that FTX published the fund's balance on its website and promoted it widely. Then: "However, according to Wang, the number shown on the website was falsified." The claim belongs to a witness, and she says so.
Second, the exhibits. She shows the tweet the prosecution used, labelled with its exhibit number, and a short exchange from the courtroom in which the witness answers that the number was fake and the real one lower. Then she shows the code itself, with its exhibit number, and points to the line where the balance change was calculated.
Third, her interpretation, marked as hers. "This is pretty damning." It is short, and it comes after the evidence.
Fourth, the strongest innocent explanation. Before concluding, she considers the best case for the other side: an inaccurate number could be explained as an old figure someone forgot to update. Then she explains why the code does not fit that story: the fund was being increased by a random number that was not tied to any real balance.
Fifth, the verdict, in the conditional. "That would be fraud."
The word "would" matters. The trial was still running. In this passage she does not state a legal finding; she says what the documented behaviour would amount to. Her headline is blunter ("The fraud was in the code"), so the conditional belongs to the argument rather than to the whole post. The reader has seen every step that leads to the verdict and can disagree with any one of them.
Four layers, kept apart
The study found that White separates fact, interpretation and opinion by position and form [2]. A fact appears as a plain statement with a link or note. An interpretation is introduced as one. An opinion arrives last, often as a short ironic remark. In practice that gives four layers.
The event, with a number and a source. Her recap issues list incidents as single lines: an actor, an action, an amount and a link, such as "Goledo Finance hacked for $1.7 million" [4]. In longer stories she gives the exact figure instead of an adjective: "He paid a $50 million fine, and his company paid another $4.3 billion in penalties." [5]
The attribution. Claims that are not yet established belong to someone. In the same list as the Goledo line, an item about hacks reads "SIM swappers charged over hacks, reportedly including FTX" [4]. "Charged" and "reportedly" tell the reader how far the claim has been established.
The interpretation, labelled. In her report from the seventh day of the trial she writes: "The picture that I think the defense was trying to paint is this:" and then sets out the defence's case before assessing it. Her assessment begins with her own doubt: "The problem is that I'm not sure this point was made terribly effectively." [6]
The verdict, last and short. Only after these layers does the judgement appear, and it is usually brief. Because the layers are visible, the reader can tell where the documents end and her view begins.
Notes as a second layer
Footnotes let her keep the body readable while each claim keeps its address. In Issue 81, a single sentence about the $TRUMP memecoin carries a note number: "Trump has already extracted around $350 million from crypto buyers via his $TRUMP token." [7] At the bottom of that issue sit more than two dozen numbered notes pointing to court filings, government statements and news reports.
This is a reader-facing choice, not decoration. A Nieman Lab profile of her newsletter observed that most of its articles "have generous footnotes and references" [8]. The reader who wants to stop at the sentence can; the reader who wants to check it is one click away from the source.
She also discloses her own position. Many issues end with a short line stating that she has disclosures about her work on cryptocurrency; the study found this line in 8 of the 26 texts [2]. A critic who asks for accountability from others also points readers to her own interests.
Fairness that can be checked
Sourced criticism also means saying what you do not dispute. In a 2024 essay on AI, she writes: "Some are surprised when they discover I don't think blockchains are useless, either." [9] She then explains what blockchains are designed to do well and what they trade away. The concession makes the rest of the criticism harder to dismiss as reflex.
When she challenges a claim, she tests it against dates. In a 2022 post on the argument that the industry is still "early", she lists when things launched: Mt. Gox in 2010, stablecoins and NFTs in 2014, a well-known DAO in 2016, popular smart contracts in 2017. Only then does she ask: "How long can it possibly be 'early days'?" [10] The question is sharp, but the dates do the arguing.
What does not transfer
Some of her tools belong to her and should not be copied. Her irony and her sharpest vocabulary are personal. So are the names of her projects and her standard formulas. Our study's recommendations for our own writing take from her the sourced incident line, attribution through documents, conflict disclosure, the short verdict and the honest concession, and leave the polemic behind [2]. Words like "fraud" belong in our texts only when a court has used them, or when the conditional form and the full evidence are both on the page.
A worked example
The protocol, the figures and the addresses below are invented for illustration.
A lending protocol, call it Harbor, says on its homepage that deposits are protected by a $50 million insurance fund. A reader asks whether that is true.
Version A, criticism without a chain. "Harbor's insurance fund is a fiction. The team is misleading depositors, and anyone who trusts them is going to get burned."
It may even be right. But the reader cannot check any part of it and cannot tell a researched view from a grudge.
Version B, criticism linked to sources. "Harbor's homepage states a $50 million insurance fund (screenshot, 10 November). The fund address named in its documentation held $3.2 million in stablecoins at 12:00 UTC that day (block explorer link). The documentation says the rest is held off-chain, but it names no custodian and links to no attestation. One explanation is a recent payout that the page has not caught up with; but the explorer history shows the address holding between $3 million and $4 million for the whole of the past 90 days. Until Harbor publishes proof of the off-chain funds, the $50 million figure cannot be verified, and depositors should treat the on-chain $3.2 million as the cover they can see."
Version B is harsher in substance and calmer in tone. Every number has an address. The innocent explanation is stated and tested. The conclusion is limited to what the evidence supports.

A method for your own criticism
- Write the verdict last. Draft the evidence first. If the verdict does not follow from what is on the page, change the verdict, not the evidence.
- Give every number an address. A link, a note or a named document, with a date for anything that can change.
- Attribute what is not established. "According to the filing", "the regulator alleges", "reportedly". Say how far each claim has travelled.
- Label your interpretation. "My reading is", "in effect", "I'm not sure". The reader should see where the documents end.
- State the best innocent explanation and test it. If you cannot think of one, you have probably not understood the case yet.
- Concede what you do not dispute. One honest concession is a signal the reader can check.
- Disclose your own interests. Before asking others to be accountable.
Common mistakes and how to fix them
1. The link as decoration. A link sits after a claim but the page it leads to does not say that. Fix: open every source and find the sentence that supports your claim.
2. Sources for the facts, none for the verdict. The events are documented, then the conclusion jumps beyond them. Fix: check that each part of the verdict rests on something shown above it.
3. Legal words without a legal finding. "Fraud", "theft", "scam" stated as facts. Fix: use the court's words when there is a ruling; otherwise describe the behaviour and use the conditional.
4. No room for the other side. The strongest innocent explanation is missing, so the reader suspects it was avoided. Fix: state it in its best form and show why the evidence does or does not fit.
5. Tone doing the work. Sarcasm replaces the missing document. Fix: remove the irony and see whether the paragraph still convinces. If not, it needs evidence, not style.
What this article cannot tell you
The measured part covers three writers in one genre, compared with each other. The counts are automatic, and the coding of attribution was done by rules and machine readings rather than two independent human coders; the study rates its own reliability as class B [2]. Part of the third-party speech in her court reports entered the counts as her own text. The passages discussed here were chosen to show the technique, so they show her method at its clearest, not on average. We did not test whether readers trust sourced criticism more; this article describes a practice and argues for it.
How this article was made
| Task | What the AI did | What a person did | How it was checked |
|---|---|---|---|
| Research | Collected the corpus of 78 public texts, counted numbers, links, notes and other features, wrote study notes for each writer | Chose the questions and the writers to study | Figures compared with the study's metrics file [1]; every quoted sentence found on the original page [3]–[10]; publication dates taken from page metadata |
| Draft | Wrote the first draft from the study notes | Edited the text and the structure | A separate review compared figures and quotes with the sources |
| 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. The author has no relationship with Molly White or the publications discussed; the protocol in the example is fictional.
Responsible for this article: Dmitriy Lastovka, author. Corrections: contact form.
Sources
Labels: study (our own analysis), primary text (the writer's original page), profile.
- Dapp.Expert style study, pilot of three writers: comparison and metrics (corpus of 78 texts, 160,115 authorial words, of which 63,712 by White; metrics file generated by the study's script, 4 November 2026). Study (internal, not peer-reviewed).
- Dapp.Expert style study, Molly White: work with fact, number and source; techniques; reliability (class B); third independent coding pass, 4 November 2026. Study (internal).
- Molly White, "The FTX trial, day four: The fraud was in the code", Citation Needed, 9 October 2023. Primary text. https://www.citationneeded.news/the-fraud-was-in-the-code/
- Molly White, "Issue 50 – Bitcoin busts", Citation Needed, 2 February 2024. Primary text. https://www.citationneeded.news/issue-50/
- Molly White, "Issue 95 – The pardon was the payoff", Citation Needed, 28 October 2025. Primary text. https://www.citationneeded.news/issue-95/
- Molly White, "The FTX trial, day seven: Sam, I guess", Citation Needed, 13 October 2023. Primary text. https://www.citationneeded.news/ftx-trial-day-seven-sam-i-guess/
- Molly White, "Issue 81 – Crypto crime is legal", Citation Needed, 8 April 2025. Primary text. https://www.citationneeded.news/issue-81/
- Sarah Scire, "Independent journalist Molly White knows how to follow the memecoin", Nieman Lab, 14 August 2025. Profile. https://www.niemanlab.org/2025/08/independent-journalist-molly-white-knows-how-to-follow-the-memecoin/
- Molly White, "AI isn't useless. But is it worth it?", Citation Needed, 17 April 2024. Primary text. https://www.citationneeded.news/ai-isnt-useless/
- Molly White, "It's not still the early days", personal blog, 14 January 2022. Primary text. https://blog.mollywhite.net/its-not-still-the-early-days/










