AI tools helped gather research and write the first draft of this article. I edited it and checked every quote against the original essays on ben-evans.com. The writing-style figures come from an internal study prepared for Dapp.Expert with AI research tools; it is not peer-reviewed. The cover illustration is AI-generated. Details at the end.
In July 2026, Benedict Evans published a long essay on how the price of AI tokens might settle. Along the way he compared the AI build-out with three older industries: fibre networks in the dot-com bubble, mobile data and semiconductor manufacturing. Then he wrote a sentence that seems to undercut all three: "analogies don't have predictive value" [1].
That is not a contradiction. It is the method. For Evans, a comparison is not evidence that something will happen. It is a tool for finding the right question, and every comparison comes with a note on where it stops working. That habit is worth learning, because comparisons are everywhere in writing about AI and crypto, and most of them are used as proof.
This is the ninth 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 builds explanations around experiments. This one looks at the second writer in our study and his main tool: the comparison.
Why Evans, and why comparisons
Dapp.Expert's internal study of three writers who cover AI and technology, Ethan Mollick, Benedict Evans and Molly White, coded short passages from their texts for several features [2]. Analogy was Evans's most distinctive one: it appeared in 39% of his sampled passages, against 20% for Mollick and 13% for White. He also asks the most questions (13.7% of his sentences) and uses the fewest links (3.1 per 1,000 words). The study's summary is that Mollick argues from his own experiments, White from documents, and Evans from reasoning and comparison.
(Internal research, not peer-reviewed. The percentages come from one close reading, by an AI coder rather than a person, of about one in ten passages (51 passages for Evans). Analogy was among the hardest features for the automatic rules to catch: their agreement with that reading was low. A second, blind machine reading found analogy in 33% of Evans's passages, against 39% in the first. Treat the figures as a direction, not a measurement.)

The few links matter for what follows. A writer who relies on reasoning more than on cited documents needs another way to keep the reasoning honest. In Evans's essays, much of that work is done by how the comparisons are chosen and limited.
Job 1: sorting cases by what happens next
In a 2017 essay on how people dismiss new technology, Evans starts from a common argument: important inventions often looked like toys at first, so this new toy might be important too. He points out that both this defence and the dismissal it answers ("that's just a toy") have "no predictive value": many things that looked like toys stayed toys [3].
Then he uses three comparisons to replace it with a better question. The Wright Flyer of 1903 was small and flimsy, but it had a clear path of improvement: better materials, bigger engines. Car phones of the late 1940s had no such path, because shrinking them needed the whole computer revolution and serving millions of callers needed new radio science. "In 1947, mobile phones were a toy," he writes. Rocket packs, the third case, carried one person a few hundred metres and could not be improved, because more fuel means more weight.
The three cases look alike on the surface: each one is a small, impractical machine. The comparison is useful because it separates them by something else. The question he draws from them is whether a technology has a roadmap: "Do you know what to do next?" [3]
What to take from this: a comparison of several cases can sort them by the feature that decides the outcome. It works best when the cases look similar and end differently.
Job 2: carrying a mechanism from the past
In a 2023 essay on AI and jobs, Evans uses a nineteenth-century case to explain an economic mechanism. "In the 19th century the British navy ran on coal," he begins, and engineers expected that more efficient steam engines would mean less coal. The economist William Stanley Jevons argued the opposite: engines that were cheaper to run would be used more, and for new things, so total coal use would rise [4].
Evans then follows the same mechanism through office history: typewriters, adding machines, mainframes and spreadsheets made clerical and accounting work cheaper, and employment in those jobs went up. A spreadsheet, in this telling, did not remove accountants; it changed what they did.
The important part comes next. He does not stop at the precedent. He lists two counter-arguments to it, that AI may move much faster than earlier waves and that it may be far more general, and answers each one separately. The precedent sets up a question; the counter-arguments test it.
What to take from this: a historical case is most useful when it carries a mechanism you can name ("cheaper means more use"), not just a mood ("people worried before, too"). Then say why this time might be different.
Job 3: choosing the comparison by the mechanism that matters
The token-pricing essay shows the most careful use of all [1]. Many commentators compare today's AI data centres with the fibre networks built during the dot-com bubble, which were overbuilt. Evans checks that comparison on the point that matters and rejects it: fibre construction "was mostly fixed cost (digging holes) rather than marginal cost", whereas more AI usage means buying more compute.
So he picks a different comparison, mobile data, which he calls "a more fruitful comparison" because mobile networks also have marginal cost for capacity and also saw usage outrun capacity about 15 years ago. He then adds semiconductor manufacturing, where rising costs at the cutting edge shrank the number of leading manufacturers to a handful and then mostly to one.
Each comparison is chosen to light up one question: will this be a low-margin infrastructure business, and will it get so expensive that only a few companies can compete? None of them is offered as the answer.
What to take from this: before you use a comparison, name the one feature it shares with your subject that matters for your question. If the feature does not match, change the comparison, not the question.
Job 4: replacing a misleading frame
Sometimes the comparison is there to remove a wrong picture. In a 2019 essay on bias in machine learning, Evans argues that the term "artificial intelligence" suggests something like a thinking mind. Instead, "it's much more useful to compare them to, say, a washing machine," he writes: very good at one job, and with no idea what it is washing [5]. In the same essay he compares machine-learning bias with the problems of early databases, where people trusted whatever the system said, and concludes that this "is not, particularly, an AI problem".
A 2023 essay uses the same move in a shorter form. A lawyer who asked an AI chatbot for legal precedents got answers that looked like precedents but were not. Evans's one-line verdict on the chatbot: "It's not a database" [4].
What to take from this: a plain, everyday comparison can correct a frame that a grand one creates. The test is whether it makes the reader expect the right kind of failure.
The rule he applies to his own comparisons
Evans states the limit directly. In the token-pricing essay, after listing several more possible comparisons, he writes: "You can't prove whether something will have the same outcome as mobile by arguing how much it's like mobile" [1]. His example is the old argument that Android would beat the iPhone because it was "open", as open Windows PCs had beaten the closed Mac.
What the comparisons can show, he says, is a range: that something can be very important, very expensive and full of advanced science, and still end up with very different outcomes. And when the honest answer is unknown, he says so: "I don't think anyone can actually know the answer yet" [1].
That pairing is the core of the method. The comparison opens the question; the limit stops it from turning into a forecast.
A comparison check before you publish
| Question | What a good answer looks like |
|---|---|
| 1. What job is the comparison doing? | Sorting cases, carrying a mechanism, choosing between comparisons, or correcting a misleading picture |
| 2. Which one feature does it share with my subject? | A named mechanism, such as fixed cost against marginal cost |
| 3. Where does it break? | At least one stated difference |
| 4. Does it predict anything? | No. If the text implies it does, rewrite the sentence |
| 5. Can the reader check it? | Dates and sources for the historical case |

A worked example
The network and the figures below are invented for illustration.
Suppose you are writing about a new network whose transaction fees jumped fivefold in a week after a popular game launched on it.
The default draft. "This is the network's Amazon moment. Just as Amazon grew through years of losses, today's fee spike shows the demand that will make this chain the winner of the next cycle."
The comparison is doing the work of proof. It shares nothing specific with the subject, it states no difference, and it points to an investment-style conclusion that no comparison can support.
The same paragraph after the check. "Fees rose about fivefold in the week after the game launched, according to the network's public fee dashboard. A useful comparison is mobile data around 2010: in both cases, usage grew faster than capacity, and each extra unit of capacity costs more to provide. The comparison stops there. Mobile operators could build new towers over years; this network's capacity depends on a software upgrade with no announced date. What happens to fees next is not known."
The second version uses the comparison to explain one mechanism, says where it breaks and makes no forecast. It is also longer, and that is fine: a comparison is only worth the space if it carries its own limits.
Common mistakes and how to fix them
1. Using a comparison as a prediction. "X is the next Y" is a forecast dressed as an explanation. Fix: say which feature they share and stop there.
2. Comparing on looks. Two things that look alike can work differently, as the Flyer and the rocket pack did. Fix: compare on the mechanism that decides the outcome.
3. Leaving out the difference. A comparison without a stated limit invites the reader to fill in the rest. Fix: add one sentence that begins "The comparison stops at…" or "Unlike…".
4. An old case without a date. "Like the early internet" can mean anything from 1969 to 2005. Fix: give a year and a source the reader can open.
5. Borrowing the writer instead of the method. Evans's essay titles, coined phrases and favourite images are his. Fix: take the method, one named feature and one stated limit, and use your own cases.
What this article cannot tell you
The figures on how often each writer uses analogy come from an internal, machine-assisted study of three writers, with no human second coder, and the two machine readings put analogy in Evans's passages at 39% and 33%. The examples in this article were chosen by hand from a few of Evans's open essays; his paid newsletter and his presentations were not studied. Nothing here shows that comparisons make writing more persuasive or more accurate for readers; that has not been tested. And this article cannot tell you which comparison suits your topic; it can only show how to check one.
How this article was made
| Task | What the AI did | What a person did | How it was checked |
|---|---|---|---|
| Research | Ran the internal three-writer study; collected candidate examples from Evans's essays | Chose the question and the examples to rely on | Every quote [1], [3]–[5] checked against the original essay on ben-evans.com; study figures taken from the study files [2] |
| Draft | Wrote the first draft from the notes | Edited the text, the structure and the worked example | A separate review compared figures and quotes with the sources |
| Images | Generated the cover illustration | Chose the illustration; designed the explanatory 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 commercial relationship with Benedict Evans or the other writers named; the network in the example is fictional.
Responsible for this article: Dmitriy Lastovka, author. Corrections: contact form.
Sources
Labels: primary text (the writer's own page), internal research (not peer-reviewed).
- Benedict Evans, "Ways to think about token pricing" (9 July 2026). https://www.ben-evans.com/benedictevans/2026/7/9/ways-to-think-about-token-pricing Primary text.
- Dapp.Expert, "Author style pilot: Mollick, Evans, White" (internal corpus study, November 2026; 24, 28 and 26 texts; manual coding of sampled passages). Internal research.
- Benedict Evans, "Not even wrong - ways to dismiss technology" (May 2017). https://www.ben-evans.com/benedictevans/2017/5/24/not-even-wrong-ways-to-dismiss-technology Primary text.
- Benedict Evans, "AI and the automation of work" (2 July 2023). https://www.ben-evans.com/benedictevans/2023/7/2/working-with-ai Primary text.
- Benedict Evans, "Notes on AI Bias" (15 April 2019). https://www.ben-evans.com/benedictevans/2019/4/15/notes-on-ai-bias Primary text.










