Y Combinator Co-founder Paul Graham Says AI Will Eliminate Lengthy Writing
Y Combinator co-founder Paul Graham stated that AI is particularly lethal to writing that takes too long to get to the point: there will always be shorter alternative texts, and 99% of readers will choose the shorter version.
This aligns with his consistent view on writing. He advocates writing like speaking, reading aloud, and then removing any sentences that do not sound conversational; he also mentioned that he writes articles to the point of being nearly impossible to cut, and summarizing often loses key points. In his 2025 article "Writes and Write-Nots," he predicts that as schools and workplaces hand writing tasks over to models, fewer people will write, separating those who can write from those who cannot, while "writing equals thinking."
He also expressed disdain for another type of AI tone. This year, he received numerous emails from founders that suddenly turned into a "hard news style," which he judged to be written by models, as the founders had never written this way before; once he recognizes it as AI, it becomes difficult to read seriously. He also predicts that the term "writer bottleneck" will disappear: 99% of people will let models produce first drafts after a few days of struggle.
In contrast to the tradition of long articles, magazine editors still require several thousand-word reports to place the sugar at the end, allowing readers to walk a distance. Generative models are tuned to excel at summarizing and keeping every sentence "on point," making short posts and memos popular, while novel-like delayed revelations become harder. The term "work junk writing" has emerged on the corporate side: models can write long but not accurately. He views programming as the same logic—previously limited by how much valuable text a person could produce, AI has now exploded in coding.
Reader attention is thus being repriced. Emails, memos, and fundraising materials must prove within the first few seconds that they are not fluff; authors willing to revise each sentence fifty times treat density itself as a barrier against summarization. On the other extreme is complete outsourcing of thought, retaining only the send button.
In market mechanisms, the buyers are time-strapped readers and investors, while the sellers are secondary texts that can be generated infinitely. Funding is shifting from content factories that pay by word count to those that charge based on "whether it can be skipped." Beneficiaries are authors and product copywriters who can compress thoughts into non-removable paragraphs; those under pressure are media and PR drafts that rely on padding, background, and parallelism to fill space. The event-driven nature comes from models making "shorter alternatives" the default option.
Additionally, he did not provide a sample for the 99% figure, nor did he differentiate the speed at which literary, academic, and business styles are being replaced.
Source: Public Information
ABAB AI Insight
Graham has viewed writing as a startup filter for twenty years: if the YC application letter is unclear, the company is also unclear. After the emergence of AI, he first warned against outsourcing thought, then against style forgery, and now adds a third critique—length itself becomes the original sin. This is not a new aesthetic; it is his push of "Write Like You Talk" to the distribution level: platforms and timelines have already rewarded brevity, and models have made short texts an infinite supply. The historical parallel is the inflation of fluff after printing reduced transcription costs, and the blog era turning long articles into symbols of subscription status; this time, in reverse, generative costs have dropped to near zero, leading to the demise of long, non-pointed texts first.
Capital has already adjusted accordingly. Search and information streams slice dwell time, consulting firms use models to produce fifty pages and then manually write one page of conclusions, and PR drafts compete on who gets summarized first. Graham's own hedge is to write articles in a way that cannot be losslessly compressed, equating to using information density to counter automatic summarization. Y Combinator applies the same standard to founders: if an email resembles a news style, trust discounts immediately appear.
Analogies include Hemingway being trained for compiler-like conciseness, Stripe documentation writing interface descriptions to the point of being unable to shorten further, and Bloomberg terminals compressing the world into headlines. The industry is transitioning from "content production expansion" to "attention quota control." Long reports will not disappear, but each paragraph must prove it contains judgments that cannot be replaced by models.
Structural judgment belongs to technical substitution. The mechanism is: when shorter texts can be generated instantly, the reader's opportunity cost shifts from "finishing an article" to "why not read the shorter one." Thus, the pricing power of writing shifts from the author's elaboration rights to the reader's skipping rights. Fluff is not an aesthetic issue; it is competing with zero marginal cost substitutes.
ABAB News · Cognitive Law
When key points are delayed, readers will not wait.
The longer it is written, the more it resembles replaceable inventory.
Models lower the cost of word count while raising the threshold for being read to completion.