Anyone who runs a newsletter has tried using ChatGPT to draft it. The output usually fails the same way: it's grammatically perfect, structurally generic, and reads exactly like AI wrote it. Subscribers can tell. Open rates drop. The newsletter dies slowly.
The answer isn't to avoid AI — it's to use AI for the slow parts (research, outlining, link gathering) and keep the human voice for the parts that matter (opening, closing, opinions, jokes).
What AI is genuinely good at
- Research — gathering relevant articles, papers, news from the past week
- Outlining — organizing 10 raw sources into a coherent structure
- First drafts of summary sections — neutral overview blurbs
- Subject line A/B candidates — generating 10 options is faster than writing 1
- Spell/grammar check — already obvious
- Tone matching past issues — given enough examples
What AI is bad at
- Your specific voice — even with examples, AI flattens to a generic "insightful newsletter" voice
- Strong opinions — AI hedges; subscribers don't subscribe for hedges
- Jokes — AI humor is mostly puns and groans
- Personal stories — AI invents details that aren't yours
- Calling out specific people or companies — AI is over-cautious in a way that kills sharpness
- The opening hook — generic openers ("This week was busy in AI...") kill open-to-click rates
A workable weekly workflow
Monday: research dump. Use Perplexity, Felo, or just Claude with web access to gather. Prompt: "What were the 5 most important developments in [your beat] this past week? Include source links and a one-paragraph summary of each. Skip funding announcements unless they're material."
Tuesday: outline. Take the research. Decide what's actually interesting (not the same as "important"). Cut to 3 stories, plus your other regular sections (links, recommendation, etc).
Wednesday: draft summaries with AI. For each story you'll cover, give Claude or GPT the source links plus your angle. Prompt: "Draft a 150-word summary of [topic] with this angle: [your take]. Match the tone of these previous issues: [paste 2-3 issues]. Keep voice casual, opinionated, no hedging."
Thursday: write the parts that matter yourself. The opening, your specific opinions on what matters, the closing, any jokes, anything personal. AI can copy-edit but shouldn't first-draft.
Friday: review, fix, schedule. Read the whole thing aloud. Anywhere you stumble, rewrite. Anywhere it sounds AI-generic, rewrite.
Total time: 4-6 hours per issue, vs 8-12 without AI assistance.
The voice-matching trick
Give AI 5-10 of your previous issues as examples in context. Ask it to write in that voice. This works better than just describing your voice ("casual, opinionated") — examples are denser signal.
But: AI won't perfectly match. It will catch ~70% of your style markers. The remaining 30% is what makes you you. Always edit the AI draft for voice; never copy-paste straight to publish.
What never to outsource to AI
- The hook. The first sentence determines whether someone reads or deletes. Write it yourself.
- The recommendation. If you recommend a tool / book / article you supposedly used, AI won't have actually used it. Write only from real experience.
- The take. What you specifically think about a story. AI can give you balanced overview; you must add the angle.
- Calls to action. Any "reply with your thoughts" or "forward to a friend" should be in your voice, not generic.
- The closing. Same logic as hook — last impression determines unsubscribe rate.
When NOT to use AI for your newsletter
If voice is your differentiation. Substacks like Sinocism, Stratechery, and Pivot succeed because the voice is the product. Adding AI to those would be like running an Anthony Bourdain show but having ChatGPT write the narration.
If you're early in finding your voice. Newsletter writers in their first 20 issues are still discovering what they sound like. AI assist makes this slower because you're editing AI text rather than discovering your own.
If your audience is sophisticated and pattern-matches AI tells. Tech-literate readers can spot AI-written content within a paragraph. If your audience is engineers, designers, writers, the AI tax on trust is high.
Subscription growth and AI assistance
There's a real correlation between newsletters that admit AI assistance and lower subscriber engagement, vs newsletters that hide AI assistance entirely. The honest middle path: use AI in your private workflow, never publish AI-drafted text without substantial editing, and don't lie about it if asked.
Most successful AI-assisted newsletters disclose the workflow honestly when readers ask. They don't lead with it.
A specific prompt template
For each story you cover:
Draft a section for my newsletter on [topic].
Context:
- My audience: [who reads it]
- My angle on this: [your specific take]
- Tone: casual, opinionated, slightly skeptical, never breathless
- Length: 150-250 words
- Don't end with a question to readers
- Don't use the words "unprecedented," "game-changing," "revolutionary"
- Match the voice in these examples: [paste 3 past issue blurbs]
Source material:
[paste articles or links]
The "banned word" list is critical. AI defaults to marketing language. Banning it forces non-generic output.
Decision tree
- Newsletter where voice IS the product: AI for research only, write everything else
- Newsletter where you cover topics with consistent format: AI for summaries + own writing for hooks / opinions
- High-volume daily news: AI for first drafts + human edit + always disclose
- New newsletter, finding your voice: Don't use AI for content, only for research
Next steps
- Read about prompt engineering for matching tone
- Try newsletter-specific tools: Beehiiv AI, Substack notes
- Build a personal style guide you paste into your AI prompts
- Track open and click rates before/after AI integration to see if it's helping