Here’s a slightly awkward admission from someone who sells an AI marketing product: I’m tired of AI content. I scroll LinkedIn in the morning and I can feel it within a line. The opener that could start any post. The tidy list of three. The closing question nobody was ever going to answer. I don’t finish those posts anymore, and I’ve stopped feeling bad about it.

I’m clearly not the only one. But I think most of the industry has taken the wrong lesson from the backlash, and it’s a lesson I very nearly built a company on.

How much of it there is now

Last December, Merriam-Webster made “slop” its word of the year, defining it as low-quality digital content churned out in bulk by AI. When the dictionary gives something a name, it has stopped being a niche complaint.

The volume behind the word is real. Graphite, an SEO firm, has been sampling English-language articles from across the web since 2020. Twelve months after ChatGPT launched, 36% of new articles were primarily AI-generated. By early 2025 it was about half, and it has sat at roughly 50% ever since.

LinkedIn is further along. Originality.ai, which sells an AI detector (so read their numbers with that in mind), checked 5,000 public LinkedIn posts of 100+ words from July this year and classified 81% as likely AI. In late 2024 their figure was around half. In the same month, LinkedIn started rolling out a “Seems like AI slop” option in the post menu, and its chief product officer called slop a top priority. When a platform builds a button for something, you know how its members feel about it.

What readers do about it

Readers have responded the only way they can. They discount everything. Raptive, which sells advertising for publishers, surveyed 3,000 US adults and found trust dropped by nearly half when people thought an article was AI-generated, whether or not it actually was. Purchase consideration for products advertised next to it fell 14%.

Sit with “whether or not it actually was” for a second. Suspicion is now the default setting. A careful, human post that happens to share the rhythm of generated text pays the same penalty as the real thing. So slop taxes everyone writing in the same feed, including the people who never touched a generator.

And it isn’t even paying off for the people producing it. Orbit Media’s annual survey of content marketers shows AI adoption going from 65% to 95% in two years. Yet in their early 2025 data, marketers who used AI to write complete articles were less likely than average to report strong results. Everyone got the tool. The ones who handed it the whole job did worse.

The mistake: building the tool to replace you

This is where I think we’ve gone wrong as an industry, and I include myself.

Almost every AI content product is designed around one promise: you don’t have to do this anymore. Type a topic, get a post. Paste your URL, get a month of content. The pitch is replacement, and it’s a very attractive pitch to a founder who hates writing and doesn’t have a marketer.

The trouble is what gets replaced. A post that works on LinkedIn works because of what’s in it: the customer conversation that changed your mind, a number from your own pipeline, a position your competitor wouldn’t dare take. That material lives in exactly one place, which is you. A tool designed to replace you has cut itself off from the only source of the thing readers are actually looking for. So it fills the gap with the average of everything it was trained on, and the average of everything is slop. The tool isn’t malfunctioning when it produces it. It’s doing precisely what it was built to do with the inputs it was built to have.

That’s also why better models haven’t fixed any of this. Fluency was never the missing ingredient. You can make the average smoother and smoother, and it’s still the average.

Where I got this wrong first

When I started Groundworks AI, our early thinking leaned the same way. Describe your business once and the system takes it from there. And the drafts were competent. They read fine. Most mornings I’d approve one and think, yeah, that’s fine. But fine isn’t a reason for anyone to stop scrolling.

It took me longer than I’d like to admit to see that competence was the problem. The drafts were good at sounding like a founder. They just didn’t sound like me, because nothing in them could only have come from our company, our customers or our week.

The principle: the tool has to work to get to know you

So we flipped the design brief. The question stopped being “how do we take this off your plate?” and became “how does the tool earn the right to write as you?”

Where we landed is that the tool needs to work to get to know you, and keep working at it. The way a good new colleague does in their first months, and then keeps doing long after. For us that comes down to two things.

It learns from what you change. Every edit, every skipped draft, every “I would never say that” is information about you, and it should make the next draft need less correcting. If you’re fixing the same thing in September that you fixed in July, the tool isn’t learning anything. It’s just generating.

It goes looking for your substance instead of inventing it. The material that makes a post worth reading (what happened, what you think, what you’d bet on) has to come from you and from your world: what your company actually did this week, the sources you actually follow. The tool’s job is to find that and bring it forward with as little effort from you as possible. Its job is never to make it up.

That second point matters more than it looks. The difference between a tool that replaces you and one that knows you isn’t the quality of the prose. It’s where the content comes from.

Your part in this is small, but it isn’t zero: read, react, correct, approve. Minutes, not hours. I’d be wary of anyone promising you zero, because zero is exactly how you end up with a feed full of posts nobody wrote.

What I’m still unsure about

We haven’t solved this, and I don’t want to pretend otherwise. Getting substance out of a busy founder without turning it into homework is hard. The line between “asks enough to know you” and “asks so much it becomes a chore” is something we’re still tuning every week. And there’s a fair question about whether any amount of learning closes the gap completely, or whether the best posts will always be the ones you sat down and wrote yourself on a Tuesday night. I suspect some of them will be. I’m fine with that.

What I am sure about is the direction. What readers are rejecting is content with nobody behind it, and that’s a problem you can design your way out of. Build tools that replace the person and you’ll make more slop, faster. Build tools that get to know the person and you at least have a shot at the opposite.

Obviously I have skin in the game here. Scout is our attempt at the second kind of tool, so weigh my view accordingly. That’s context, not a pitch. I’d be making the same argument if I were the one buying.

If you’ve found an AI setup that genuinely sounds like you, or you’ve given up on them altogether, I’d really like to hear why. Drop me a note.

Sources