Divya van Mahajan

5 Rules for Better AI Writing

ai-daily-brief-podcast
Listen to episode →

5 Rules for Better AI Writing

Overview

This talk is an episode of The AI Daily Brief, a daily podcast and video covering important news and discussions in AI. The episode uses a recent controversy — investor Stanley Druckenmiller’s AI-written op-ed published in the Wall Street Journal — as a jumping-off point to develop a practical framework for when, how, and how much to rely on AI for different types of writing. The central argument is that AI writing is now a permanent fixture of professional life, but its normalisation does not eliminate the obligation to produce quality output; different writing contexts demand different levels of AI involvement.

Source video URL: Not available (title: 2026-08-26-5-rules-for-better-ai-writing)

Prerequisites

  • Basic familiarity with large language models (LLMs) and tools such as ChatGPT and Claude
  • General awareness of how AI text-generation works and what common “AI-isms” (recognisable stylistic patterns) look like
  • Understanding of professional writing contexts: emails, memos, social media copy, op-eds, and marketing copy
  • Awareness of AI detection tools (e.g., Pangram) is helpful but not required

Main Points

The Druckenmiller Controversy as Catalyst

  • Investor Stanley Druckenmiller published an op-ed, “Let the Bond Market Speak,” in the Wall Street Journal critiquing Treasury Secretary Scott Bessent and current monetary policy.
  • The piece was widely identified as AI-generated — AI detectors rated it 100% likely AI-written — due to an abundance of recognisable AI-isms (e.g., “It’s not X, it’s Y” constructions).
  • Rather than apologising, Druckenmiller confirmed AI use unapologetically, comparing it to using a calculator for maths; the WSJ’s opinion editor backed him, stating the relevant test is whether the argument is the author’s genuine opinion.
  • The episode situates this controversy as a pivot point: if AI writing is now openly accepted, the operative question shifts to how to do it well.

The Broader Debate: Disclosure, Legitimacy, and Effort

  • Critics argued AI writing without disclosure constitutes a form of plagiarism and undermines the premise of a thought piece (Jason Calacanis, Jesse Livermore).
  • Defenders noted that ghostwriting by human staff for high-profile bylines has been common practice for decades (Joe Wiesenthal, Sharon Goldman).
  • A key nuance emerged: the backlash was less about AI use per se and more about the perception of low effort — the AI-isms were so obvious and so easily fixable that they signalled the author had not engaged seriously with the material.
  • Balaji Srinivasan’s framing was cited: “AI doesn’t do end-to-end, it does middle-to-middle. The new bottlenecks are prompting and verifying.”

Rule 1 — Different Types of Writing, Different Types of Rules

  • An email, a strategy memo, a LinkedIn post, and an op-ed are each trying to achieve fundamentally different things; treating them identically when applying AI is a category error.
  • The appropriate level and mode of AI involvement should be calibrated to the purpose, audience, and expectations of each format.

Rule 2 — The Purity Test Will Fade; the Quality Test Will Not

  • The expectation that writing must be entirely human-produced will likely become obsolete within a few years, as multiple commentators in the debate predicted.
  • However, the quality standard will not disappear — and may actually intensify, because easier inputs raise audience expectations for outputs.
  • Accepting AI writing does not mean accepting poor writing as a consequence.

Rule 3 — Quality Is Correlated with Effort (or the Perception of Effort)

  • Readers consistently assess quality through the lens of perceived effort; obvious laziness undermines the credibility of the underlying argument, not just the writing itself.
  • Druckenmiller’s op-ed is the case study: the glaring, easily-corrected AI-isms suggested he had not invested effort in the writing, which in turn made readers doubt the depth of his thinking.
  • The lesson: even if AI generates the words, the author must invest visible effort in editing and refinement.

Rule 4 — Longer Is Not Better; Brevity Is the Goal

  • Early AI writing tools excelled at producing volume rather than precision; the result was long, “eye-bleeding” documents that said a lot without saying the right thing succinctly.
  • This echoes a principle great writers have always known, illustrated by a 17th-century Blaise Pascal quote: “I have made this longer than usual because I have not had time to make it shorter.”
  • The next generation of AI writing norms will reward concision, not length; enterprise users in particular should be aware of “workslop” — AI-generated verbosity that fills space without adding value.

Rule 5 — Writing Is Thinking; Outsourcing Writing Risks Outsourcing Thinking

  • The process of constructing a thesis, selecting supporting arguments, and crafting phrasing is itself a form of thinking; the words on the page are the output of that cognitive work, not its purpose.
  • Using AI to generate writing does not automatically mean outsourcing thinking, but it creates a constant and serious risk of doing so.
  • The speaker’s personal example: rather than prompting Claude to build the “5 Rules” episode structure, he worked through the framework manually in Notion first — a slower process, but the right one for this context.

Application to Specific Writing Contexts

  • Emails: Generally safe for AI; the prose quality is rarely judged. However, for quick interactive exchanges, dictation tools (e.g., Whisper Flow) may be more efficient than full AI drafting. Templated responses (e.g., sponsorship replies) are the clearest good use case.
  • Meeting notes/summaries: Very safe for AI because the task is compression, not original composition. Caveat: AI may summarise what was said thoroughly but miss what mattered most — a human should add a brief top-level note on the one or two most important points.
  • Internal strategy memos: Sneakily risky. AI may fill in generic advice from training data rather than reasoning from the specific organisational context. Best approach: use AI to support thinking (outlining, iteration) as a first phase, then use AI to generate the actual prose from a tight, human-constructed outline as a second, entirely separate phase.
  • Social media copy: Mixed results. AI performs better on shorter formats; longer LinkedIn essays give AI-isms more room to appear. A deeper problem: social platforms increasingly reward engagement (commenting, responding) and human signal, not just content volume. AI writing pipelines can increase output and sharing but quickly hit a ceiling. Upfront voice-tuning is the most effective lever.
  • Marketing copy: The area the speaker finds most frustrating. AI-isms are especially costly here because marketing requires distinctiveness. A specific failure mode: when given a correction (e.g., “stop presuming about the reader”), the model tends to convert the correction directly into copy (“We assume nothing about the reader”). Best use: rapid iteration for ideation — asking AI to generate 20 tagline variants to spark human selection.
  • Op-eds and persuasive essays: The highest-stakes context. The format’s entire purpose is persuasion; perceived laziness immediately undermines the argument’s force. The required approach: the author must first achieve complete clarity on thesis and supporting points (the “five-paragraph essay” discipline), then hand off to AI for drafting, then review and eliminate obvious AI-isms before publication.

Key Concepts

  • AI-isms: Recognisable stylistic patterns that signal AI-generated text, such as repetitive “it’s not X, it’s Y” constructions, excessive self-congratulatory framing, or hollow rhetorical flourishes.
  • Workslop: AI-generated content that is verbose and lengthy but lacks precision, insight, or genuine value; a portmanteau of “work” and “slop.”
  • Middle-to-middle (Balaji Srinivasan): The idea that AI handles the intermediate stages of a task well (researching, drafting, organising, tightening) but that the beginning (defining the idea and intent) and the end (evaluating quality and truth) require human judgment.
  • Purity test: The expectation — predicted to be short-lived — that legitimate writing must be entirely human-produced with no AI involvement.
  • Quality test: The enduring standard that writing, regardless of how it is produced, must meet audience expectations for clarity, precision, and evident effort.
  • Writing is thinking: The principle that the act of composing — choosing a thesis, structuring arguments, selecting phrasing — is itself a cognitive process; generating text without performing that process risks producing hollow output.
  • Pangram (AI detector): A tool used to estimate the probability that a piece of text was generated by AI; cited in the episode as having rated Druckenmiller’s op-ed at 100% likelihood of AI generation.
  • Whisper Flow: A dictation tool referenced as an alternative to AI drafting for short, interactive communications such as emails.
  • Voice tuning: The process of extensively providing an AI model with examples of on-brand or preferred writing styles in order to improve the consistency and authenticity of its output.

Summary

The speaker uses Stanley Druckenmiller’s unabashedly AI-written Wall Street Journal op-ed as a lens through which to examine a broader and more durable question: now that AI writing is openly normalised, what does it mean to do it well? The episode’s five rules establish that writing type must determine AI strategy, that quality standards will survive even as purity norms erode, that perceived effort remains a proxy for the strength of an underlying argument, that brevity is superior to AI-generated verbosity, and that the greatest risk of AI writing is inadvertently outsourcing the thinking that should precede the words. Applied across specific formats — from emails and meeting notes (relatively safe) to strategy memos and op-eds (high risk without disciplined human thinking first) to marketing copy (still largely unsolved) — the framework consistently points to the same conclusion: AI is most valuable as a tool that accelerates and refines human thought, not as a replacement for it, and laziness in the use of that tool will be visible to audiences regardless of how the text was produced.