How to use AI without making slop

About half of the articles published on the web right now are written mostly by AI. That is not a guess: Graphite ran 55,400 articles through three independent detectors and found AI-generated pieces crossed 50% of new publications in late 2025. The same research found those articles largely never appear in Google results or ChatGPT answers. Half the internet’s new words, written for an audience that isn’t coming.

The word for this got its coronation in December, when Merriam-Webster named “slop” its 2025 word of the year. The lazy conclusion is “AI content bad.” This post argues something more useful: slop isn’t caused by AI. It’s caused by shipping the first draft with nobody accountable for the result. That distinction matters because it means the difference between slop and leverage isn’t which tool you use or how much of it. It’s a short list of learnable practices, and this post is that list: for hands-on work first, then for the automations that run while you sleep.

Volume got free, and the flood got a name

Merriam-Webster’s editors define slop as “digital content of low quality that is produced usually in quantity by means of artificial intelligence,” and their examples run from off-kilter ad images to junky AI-written books to talking cats. Note what the definition actually hinges on: low quality, produced in quantity. The technology is almost incidental. What AI changed is the price of volume. Producing a thousand passable articles used to cost a newsroom; now it costs an API key. When volume gets free, the only thing separating a publisher from a spammer is whether anyone is accountable for what ships.

The scale is easy to misread, in both directions. Ahrefs analyzed 900,000 new pages in April 2025 and found 74.2% contained at least some AI-generated text, which sounds apocalyptic until you see the breakdown: only 2.5% were purely AI, while 71.7% were human-AI blends. Assistance is already the norm, and most of it is invisible and fine. The slop problem lives in a specific slice: content generated at volume, shipped unread, owned by no one. You already know its tells, because they are what you skim past every day: the averaged voice that sounds like everyone, confident paragraphs that say nothing, hedges stacked on hedges, the listicle cadence, the illustration with melted hands that nobody looked at before posting.

The model isn’t the problem; the missing owner is

The best evidence that ownership, not generation, is the dividing line comes from inside companies. In September 2025, researchers at BetterUp Labs and Stanford’s Social Media Lab named the workplace version “workslop”: AI-generated work that masquerades as good work but lacks the substance to advance the task. Their survey of 1,150 US employees found 40% had received some in the past month, and that each incident took nearly two hours to untangle. The mechanism is the important part: workslop doesn’t remove work, it relocates it. The sender saved an hour; the receiver spends two decoding, correcting, or redoing. The polish is real, the thinking is absent, and someone downstream pays the difference. The researchers connect this to the most quoted AI statistic of 2025, an MIT Media Lab finding that 95% of organizations see no measurable return on their generative AI investment. A tool that saves the sender an hour while costing the receiver two doesn’t register as productivity. It registers as motion.

Here is the twist the “I can always tell” crowd should sit with. In controlled tests, people are bad at spotting AI output itself: a study of 1,276 participants found detection of synthetic text, images, and audio hovered near coin-toss accuracy. Yet workslop recipients in the Stanford survey confidently judged their colleagues: 42% found the sender less trustworthy, about half found them less capable, a third didn’t want to work with them again. Square those findings and you get the real detection story. People can’t reliably detect AI. They can detect absence: missing context, generic examples, nothing at stake, no evidence a mind engaged with the problem. Which yields the test this whole post hangs on. Before anything ships, one question: would I sign this? Not “is it good enough to send,” but “am I willing to be personally accountable for every claim, example, and emphasis in it?” Everything that follows is that question, operationalized.

Leverage starts in the brief, not the draft

The single highest-leverage minute in any AI-assisted work happens before generation: it’s the brief. “Write a post about customer onboarding” produces the average of everything ever written about onboarding, which is by definition the thing nobody needs to read. The model has no access to the only assets that make your output worth anyone’s time: your opinions, your examples, your dead ends, your customers’ actual words. If those aren’t in the prompt, they won’t be in the draft.

A working brief states an argument, not a topic (“onboarding emails fail because they celebrate signup instead of driving the second session”). It hands over your raw material: the support ticket that sparked the idea, the numbers from your own dashboard, the stance your competitors won’t take. And it says what to avoid, because the defaults are the tells. Compare two briefs for the same email. “Write a re-engagement email for inactive users” gets you the email every inactive user already ignores. “Users who skip week two never come back; our data shows the ones who stay all imported their contacts. Write a plain, short nudge from me that only asks them to import contacts, no discount, no ‘we miss you’” gets you a draft only your company could send, because only your company knew that.

The test for a finished brief is simple: could a competitor paste it into the same model and get your piece? If yes, there is nothing of you in it yet. Our prompt-writing guide covers the mechanics; the craft rule is that taste enters at the brief or never.

Edit like an editor, not a forwarder

The second practice is a posture change: treat the draft as a verdict and you’re a forwarder; treat it as raw material and you’re an editor. AI gives you the block of marble, cheap and instantly. Carving is still your job. In practice that means a real deletion pass: cut every sentence you wouldn’t have written yourself, every hedge you don’t actually hold, every “in today’s fast-paced world” the model reached for because a million other documents did. If the edit removes nothing, it wasn’t an edit.

Voice survives the same way. Write your stories, examples, and stance first, in your own flat unpolished words, and let the model structure around them, rather than asking it to invent feeling and then trying to inject yourself afterward. Read the result aloud; rewrite whatever you stumble on, because your reader will stumble there too. This is the discipline newsletter writers who survive on voice already practice, and it’s the difference between output that sounds like you on a good day and output that sounds like everyone on an average one.

Automations need review gates sized to the stakes

Hands-on slop comes from laziness. Automation slop comes from missing gates, and it’s the more dangerous failure because it compounds unattended. A scheduled AI workflow that drafts, formats, and publishes will produce its thousandth piece with exactly the care of its first, and if nobody is reviewing, that care is zero. The fix is not “don’t automate.” It’s to put a human gate at the point where the stakes justify one, and to size the gate to what one bad output would cost.

Three rules make the gates workable at volume. First, batch-generate, human-approve: let the pipeline produce fifty drafts overnight, but publishing stays a button a person presses. Second, spot-check on a cadence: for high-volume, low-stakes streams where per-piece review is unrealistic, sample a fixed percentage weekly and treat any failure as a pipeline bug, not a one-off. Third, the kill-switch rule: an automation you can’t audit and stop is one you can’t trust, so keep logs of what shipped where, and keep the off switch one click away. The economics justifying all this are asymmetric. The gate costs minutes per week; one confidently wrong claim in front of the wrong customer costs trust that took years. And the market is already repricing ungated volume: Graphite’s finding that AI-heavy articles have plateaued near 50% since early 2025, locked out of rankings, is what it looks like when distribution channels learn to route around unowned content.

Sometimes the right amount of AI is none

An honesty box, because a post about quality bars should have one. Some writing is valuable precisely because a person labored over it: the apology, the condolence note, the performance review, the founder’s letter. Outsourcing those isn’t efficiency, it’s counterfeiting the signal the reader came for. Volume has diminishing returns too. If your plan only works at a scale where no human can own the output, the plan is slop by design, and no practice in this post rescues it.

Disclosure norms are also tightening around exactly this line. The EU AI Act’s Article 50 transparency obligations apply as of August 2, 2026: AI systems must identify themselves in conversation, deepfakes must be labeled, and AI-generated text published to inform the public on matters of public interest must be disclosed, with fines up to €15 million or 3% of worldwide turnover. Notably, the text rule carves out content that underwent human editorial review, where a person takes responsibility. The regulation is drawing the same line this post does: the unit that matters is not whether AI touched the work, but whether a human owns it. Full disclosure, practiced here: this blog is AI-assisted, from research to draft, against human-written briefs, with per-post research requirements and a review gate before publish. That is the bet the whole post makes: done openly and owned, the method holds the bar.

Ten questions before anything AI-touched ships

Everything above compresses into a checklist you can run in two minutes, on a paragraph or a pipeline. It works because every question is a proxy for the same thing: is there a human in this work, and are they willing to be seen standing behind it?

The takeaway fits in a sentence: automate the drafts, never the taste. AI made volume free, which made ownership the scarce and valuable input. The people getting genuine leverage from these tools aren’t the ones generating the most; they’re the ones whose name means something because everything under it, however it was drafted, was signed by someone who read it and meant it.

Disclaimer: This article is general information, not legal advice, and reading it creates no attorney-client relationship. Laws, regulations, and court rulings summarized here reflect sources available as of August 2026 and may have changed. Consult counsel licensed in your jurisdiction before acting on any of it.