Audio narration available on the Substack version of this post.

People often say to “write like you talk.” Paul Graham has a post titled “Write Like You Talk” where he says explicitly that written language is worse than spoken language because

1. “Written language is more complex, which makes it more work to read”
2. “It’s also more formal and distant, which gives the reader’s attention permission to drift”
3. “The complex sentences and fancy words give you, the writer, the false impression that you’re saying more than you actually are”

He gives concrete advice: “Before I publish a new essay, I read it out loud and fix everything that doesn't sound like conversation. … [If you have] writing so far removed from spoken language that it couldn't be fixed sentence by sentence … try explaining to a friend what you just wrote. Then replace the draft with what you said to your friend.”

In “Politics and the English Language,” Orwell also identifies complex and formal diction as a way to mask emptiness, “to dress up simple statements and give an air of scientific impartiality to biased judgements. … A mass of Latin words falls upon the facts like soft snow, blurring the outlines and covering up all the details. … If thought corrupts language, language can also corrupt thought.”

But Orwell doesn’t make the same distinction between spoken and written language; in fact he says that “When you are composing in a hurry—when you are dictating to a stenographer, for instance, or making a public speech—it is natural to fall into a pretentious, latinized style.” And Graham himself elsewhere says that “I'll often spend 2 weeks on an essay and reread drafts 50 times” and “I'm sure there are sentences I've read 100 times before publishing them” which is the opposite of how conversations work.

In contrast, Scott Alexander claims that it only takes him “a couple of hours” to write a post and that “I don’t really understand why it takes so many people so long to write. They seem to be able to talk instantaneously, and writing isn’t that different from speech.” But then Scott is often considered a digressive or even “astoundingly verbose” writer.

There’s debate over whether speech or writing is more “complex” at all, with scholars taking sides based on the metrics they use for complexity and the datasets they analyze. In particular, there’s debate over whether speech or writing uses more subordinate clauses. Subordinate clauses are so called because they can’t stand as independent sentences. They have a few common types:

1. The adverbial clause functions as an adverb: it modifies the sentence or some element of the sentence, e.g. by adding information about time, place, manner, reason, or condition.
   1. She arrived when the party had already started.
   2. After the parade ended, we went for a walk.
   3. The game was canceled because of the heavy rain.
2. The noun clause (or “nominal” or “content” clause) functions as a noun: usually it’s a that-clause or wh-clause.
   1. He told her (that) she was smart.
   2. I know what you did.
   3. She asked where the files were.
3. The relative clause (or “adjectival” clause) functions like an adjective: it modifies a noun or pronoun.
   1. The book that I bought yesterday is excellent.
   2. Students who study regularly tend to perform better.
   3. The house where I grew up has been sold.

The most intuitive comparison of spoken and written English is between matched narratives; ask test subjects to describe the same scene with either an oral or written narrative. The two studies that use this method found more subordination in speech.

• Beaman (1984) found more subordinate sentences in speech (18% vs 13%) and more compound sentences with subordinate clauses in speech (27% vs 18%) from a sample of women presenting spoken or written narratives.
• Prideaux (1993) found that 40 oral narratives of film clips contained more given relative clauses—i.e. relative clauses that restate known information—than written narratives of the same events (5.38 vs 1.92 per narrative), while both contained similar numbers of new relative clauses (1.75 vs 1.83), which presented new information.

Other studies compared writing and conversation without any matched pairs:

• Greenbaum and Nelson (1995) used a 90,000-word corpus called ICE-GB and found that in conversation there was subordination in 39% of clusters (roughly “sentences”), much less than in academic writing (63%) and non-academic (69%); conversation also only averaged 0.7 subordinate clauses per cluster, versus 1.2 and 1.6 per cluster for academic and non-academic writing respectively.
• Using samples from the same corpus, Fang (2006) also found speech to be less complex: 12% of spoken sentences contained adverbial clauses, versus 27% of written sentences. Adverbial clauses made up 31% of clauses in writing vs only 15% in speech.
• On the other hand, Biber and Gray (2010) used a large corpus to find 9.5 adverbial clauses per thousand words in conversation, versus only 3.5 in academic writing. A few older sources from the 1980s also advance this view that speech contains more subordinate clauses.

But subordination isn’t the only measure of complexity. Consider the following two sentences:

• "The cellular anatomy of the peripheral nervous system renders it vulnerable to injury."
• "Living in the Gulf has meant living with oil."

The first is a single clause with no subordination while the second contains two nominal clauses: “living in the Gulf” and “living with oil” function as the subject and object respectively. But a reader would say that the first is more complex because it’s a longer sentence that uses longer words and abstract technical jargon. Studies consistently find that writing is more lexically dense, i.e. it has more words that convey content relative to grammatical or functional words.

For academic writing, the other major difference is that writing uses a compressed style that uses noun phrases rather than clauses to add information. These noun phrases often leave the underlying relationship implicit; even phrases like “heart disease” don’t reveal whether the meaning is “disease caused by the heart” or “disease located in the heart” or “disease affecting the heart” if readers don’t already know. Starting in the mid-20th century, multi-noun sequences like “air flow limitations” and “plasma concentration time curve” became more common in academic, newspaper, and medical prose. The relationships between these pre-modifying nouns is left implicit. The excerpt below is from the excellent Biber and Gray (2011).

The compressed style also often uses many layers of embedding in its noun phrases: consider “the effects [[of changes [in taxonomic resolution]][on analyses [of patterns [of multivariate variation [at different spatial scales]]]] [for the highly diverse fauna [inhabiting holdfasts [of the kelp Ecklonia radiata]]]].” So the main clause of a sentence can be very simple even when the phrasal modification is very complex: consider the sentence “This may indeed be part [of the reason [for the statistical link [between schizophrenia and membership [in the lower socioeconomic classes]]]].”

Lastly, a difference between speech and writing is that face-to-face conversation uses more “metadiscourse” to lighten the cognitive load for both the speaker and listener. The speaker moves from topic to topic on the fly based on what seems natural while using verbal signposts and scaffolding to mark digressions, signal a return to the main line of thought, and keep the overall structure in view. Because listeners can’t reread or rely on visual cues like headings and paragraph breaks, they depend on these markers—“to sum up,” “we’ll come back to that,” “by the way”—along with repetition and reformulation of key points to stay oriented and to know when to shift their attention back to the argument’s central thread.

So the common advice to "write like you talk" can be underspecified. It's good to avoid pretentious and formulaic cliches that mask the absence of precise thought, and separately to avoid dense and impenetrable jargon that's hard for non-experts to understand. But it's bad to write verbose and digressive meanderings without editing them. And because this style of prose is faster to write, it can occupy a large share of posts (and more so of words) in Internet forums and discourse.