Last spring, a colleague lost her father, and I did something I'm still a little ashamed of: I asked an AI to draft the condolence message. What came back was grammatically flawless, warm in a laminated sort of way, and completely hollow. I deleted every word and sent four clumsy sentences of my own. Months later she mentioned, unprompted, that mine was one of the few that had actually helped.
That moment sits at one end of my experience with these tools. At the other end: the afternoon I restructured forty pages of product documentation in under three hours, an AI doing the heavy lifting while I steered. Same tool, wildly different verdicts. Somewhere between those two afternoons is a line, and for three years I've kept notes (an actual spreadsheet, 200-plus entries) on where it falls.
This guide is my attempt to draw that line properly, using published research rather than vibes plus my own logs where the research runs out. Fair warning: I use AI most working days, so I'm biased toward the tools. I've tried to let the data push back wherever it disagrees.
| REACH FOR AI WHEN… | KEEP IT HUMAN WHEN… |
|---|---|
✓ You're staring at a blank page for a routine document ✓ The draft exists and needs tightening or restructuring ✓ You need ten angles fast and nine can be bad ✓ You can verify everything it says in minutes | × The reader is judging you through the words × The point of the task is learning or thinking × You can't personally check the facts × Effort itself is the message (grief, apology, thanks) |
What the research actually shows
Strip away the hype and the doom, and the serious studies tell a consistent story. Headline numbers first, then the fine print, where the “skip it” half of this article lives.
-40% time on professional writing tasks (MIT, Science) | +18% graded quality in the same randomized study | 97% of content marketers using or planning AI in 2026 | 1% who publish fully AI-generated work; everyone else edits |
Five studies worth knowing before you form an opinion
| STUDY | WHAT THEY TESTED | WHAT HAPPENED |
|---|---|---|
MIT / Science, 2023 Noy & Zhang · randomized, n=453 | Professionals doing realistic occupation-specific writing tasks; half given ChatGPT | Time fell ~40%, graded quality rose ~18%. The weakest writers gained most, and work shifted from rough-drafting toward ideas and editing. |
Harvard / BCG, 2023 Dell'Acqua et al. · n=758 consultants | Consultants on tasks inside vs. outside the AI's “jagged frontier” of competence | Inside the frontier: ~40% higher quality. Outside it: AI users were roughly 19 percentage points less likely to reach correct answers. |
NBER, 2023 Brynjolfsson et al. · n=5,172 agents | Customer-support agents at a Fortune 500 firm using an AI assistant | +14% productivity overall, but +34% for the least experienced workers and near-zero gains for top performers. |
MIT Media Lab, 2025 EEG study of essay writers | Brain activity and recall in essay writers using an LLM vs. unaided | AI-assisted writers showed the weakest neural engagement; more than 80% couldn't quote a sentence from the essay they had just “written.” |
Reader-trust experiments, 2024-26 Toff & Simon; Schilke & Reimann; others | How audiences react when text is labeled as AI-generated | Labeled AI text is rated less trustworthy even when readers judge it accurate and fair, a “transparency penalty” that hits the human author's credibility too. |
Read the five together and a pattern emerges: AI reliably compresses effort on structured, verifiable writing, while its costs surface in three places: your memory of the material, your reader's trust, and errors at the edge of your competence. That pattern is the basis for every verdict below.
Where AI earns its keep [USE AI]
Every task in this table shares two properties: the output is quick to verify, and nothing personal rides on the prose itself. When both are true, the data is close to unambiguous. This is where MIT measured that 40% time saving, and where marketers report reclaiming ~6 hours a week (HubSpot, 2026).
| TASK | WHY AI EARNS IT | WHAT THE EVIDENCE SAYS |
|---|---|---|
First drafts of routine documents briefs · job posts · process docs · FAQs | The blank page is the expensive part. A mediocre draft you can attack beats a perfect draft you haven't started. | MIT found AI shifted work away from rough-drafting toward ideas and editing, exactly the trade you want. 72% of PR professionals already draft this way (Statista, 2025). |
| Editing, tightening, restructuring | AI is a tireless line editor: it will cut 30% of your words without ego, at midnight, on the fourth pass. | The fastest-growing use: marketers editing with AI doubled from 19% to 38% in one year (Siege Media + Wynter, 2026). |
| Brainstorming angles & outlines | Breadth beats depth here. Ten options in ten seconds, and the nine bad ones cost you nothing. | Ideation is the top AI use among PR professionals at 82% (Statista), and even Google's search advocates endorse AI for inspiration. |
| Summaries & meeting notes | Compression of material you already know is low-risk: you'll spot any distortion instantly. | Writing and summarizing were the largest category of work usage in OpenAI's own 2025 analysis of real conversations. |
Format conversion email → doc → slide notes → checklist | Purely mechanical transformation. The substance is yours; only the container changes. | Squarely inside the Harvard/BCG “frontier”: structured input, structured output, trivially checkable. |
| Plain-language & second-language passes | Grammar and clarity leveling is where AI most resembles a very fast, very patient editor. | Gains concentrate at the bottom: the weakest writers improved most at MIT, novices gained 34% (NBER), and 71% of students report clearer prose. |
| The common thread: you stay the expert on the content; AI handles the labor of arrangement. When those roles flip and AI supplies substance you can't judge, you've crossed into the red table below. |
Where I've learned to close the tab [SKIP IT]
None of these are moral positions. Each is here because the research shows AI measurably backfires, or because my own logs show the “time saved” gets repaid with interest: in trust, accuracy, or thinking.
| TASK | WHY IT BACKFIRES | WHAT THE EVIDENCE SAYS |
|---|---|---|
| Condolences, apologies, thank-yous | In these messages, the effort is the content. Outsourcing it deletes the only thing being communicated. | Studies consistently find AI text reads as less warm, caring, and empathetic, and the gap is widest in sensitive contexts like these. |
| Opinion pieces & thought leadership under your name | Your byline is a promise that a human thought this. Break it once and the discount applies to everything you've written. | A 2026 experiment with 312 readers found authorship alone explained 18% of trust variance. Disclosing AI help lowers trust in the human author even when quality is identical (Schilke & Reimann, 2025). |
| Anything you're trying to learn | Writing is how thinking gets debugged. Skip the struggle, skip the learning: what researchers now call “cognitive debt.” | In the MIT Media Lab EEG study, LLM-assisted writers showed the lowest brain connectivity, and over 80% couldn't quote their own essay minutes after finishing it. |
Facts you can't personally verify legal · medical · financial specifics | Models are confident precisely where you can't check them. That's the worst possible failure mode. | The Harvard/BCG “jagged frontier” result: outside their competence zone, AI-assisted professionals were ~19 points more likely to be wrong, all while feeling productive. |
| Lived experience & original reporting | The model wasn't in the room. It can only produce a plausible average of rooms other people described. | One 2026 analysis found only ~14% of top-ranking results are fully AI-generated; Google's spam policy explicitly targets scaled, experience-free content. |
| Humor and strong-voice writing | Comedy and voice work by violating expectations. A system trained to predict the likeliest next word is structurally built to meet them. | Even when readers judge AI text accurate and fair, they trust it less once the seams show (Toff & Simon), and voice is where seams show first. |
| One number worth sitting with: Ahrefs found 74% of new web pages now contain some AI-generated text, yet fully-AI pages remain a small minority of what ranks, and only 1% of marketers ship unedited AI work. The market has voted: AI-flavored sameness is abundant, and abundance is the opposite of value. |
Three cases where honest people disagree
Most writing isn't a eulogy or a meeting recap; it lives in the middle. Here's where I've landed on the three cases people ask me about most, and the line I hold in each.
Blog posts & SEO content [IT DEPENDS]
AI for research scaffolding, structure, and the editing pass; you for the experience, examples, and every claim. This hybrid is what 97% adoption actually looks like; remember, only 1% publish pure AI output. Google says it judges quality, not method, but its “scaled content abuse” policy exists because volume without value shows up in aggregate.
My line: if I removed everything AI wrote, would something worth reading remain? If no, I haven't written an article. I've formatted one.
Cover letters & personal bios [IT DEPENDS]
Draft with AI if the blank page is blocking you, but recruiters read hundreds of AI-cadenced letters a week and pattern-match instantly. The transparency-penalty research applies in full: this document's entire job is to represent you.
My line: the first sentence and closing paragraph get written by hand, every time. Those are the two places anyone reads closely.
Social posts & newsletters [IT DEPENDS]
Announcements, event details, product updates? Green light: nobody expects poetry from a shipping notification. But if the account exists to build a personal following, voice is the asset, and audiences penalize what feels machine-made even when they can't prove it.
My line: informational posts, AI-assisted. Anything meant to make someone feel they know me stays human, with all the typos that implies.
Five questions I ask before opening the tool
When I'm unsure, I run the task through five questions, in order, because the early ones veto the late ones. Thirty seconds, and it has saved me from most of my worst instincts.
1. Will the reader judge me through this text?
If yes → human-first. Reputation compounds; one discovered shortcut discounts everything after it.
2. Is the point of this task the writing, or the thinking?
If it's the thinking (learning, studying, working out a position) → skip AI. Cognitive debt is real and the interest is brutal.
3. Can I verify every factual claim in under five minutes?
If no → skip, or use AI only for structure and source every fact yourself. This is the jagged-frontier trap.
4. Would I be comfortable if the reader knew exactly how this was made?
If no → that discomfort is data. Detection tools are unreliable, but so is your luck.
5. Is this inside my competence to review properly?
If yes to all the above → use AI freely, edit ruthlessly, and take the 40% saving without guilt.
Small rules that took me too long to learn
• I write the first sentence and the last paragraph myself, always. Openings and closings carry the voice; everything in between can be scaffolded.
• Numbers, names, quotes, and citations never survive unchecked. Every factual error in my three-year log came from those four categories. Every single one.
• Ask for ten bad options, not one good one. AI is a better wide-net brainstormer than a decision-maker; the choosing is your job.
• The read-aloud test. One pass, out loud. If I can't hear myself in it, it gets rewritten. Readers sense the seams before they can name them.
• Keep a “voice file.” A dozen paragraphs of your unassisted writing. When drafts drift toward the great AI average, it's your way back.
The Final Draft
The first job is assembling information: gathering, structuring, compressing, and polishing. AI often does this faster and better, and tools such as WriteNexa can take much of that repetitive work off the writer’s plate.
The second job is being a person on the page: noticing something, caring about it, and saying it honestly. Readers look for evidence of a human on the other end. A language model can imitate that evidence, but it cannot supply it.
So I now ask one question before every task: is this assembly work or person work? The first goes to the machine. The second stays with me, including the sympathy notes.
Use the tool for everything that does not need you. That is how you protect the writing that does.
