AI Slop Isn't the Problem. AI Candy Is.

People don't like bad content. What happens when AI gets so good that we don't care if it's fake?

I was listening to Mark Zuckerberg talk about Meta’s future on the July 2026 earnings call when one phrase jumped out at me.

He was talking about how AI is changing Facebook and Instagram’s recommendation systems, making them better at understanding not just what content is about, but why it’s compelling and what people are interested in.

And then he described what comes next: “a whole new and nearly infinite universe of personalized content.”

I’d also been having a bunch of conversations lately about AI slop — the uncanny images, the hollow listicles, the LinkedIn posts that read like they were written by a toaster.

And something clicked.

I don’t think AI slop is the thing I’m most worried about.

I’m worried about AI candy.

Slop isn’t good. You don’t want more of it. But what happens when the AI content is so funny, entertaining, and made specifically for you? Maybe you know it’s synthetic, and maybe you don’t. But either way, it’s so good at giving you what you want that you eventually stop asking many questions about where it came from.

That’s more like candy. Tastes so good you want more, but you know too much isn’t good for you.

We’ve run this playbook before

Think back to the early 2010s and the rise of clickbait.

Upworthy-style curiosity-gap headlines. BuzzFeed at full velocity. “You won’t believe what happened next.”

It became almost comically easy to recognize.

But it also worked.

In 2014, Facebook started demoting clickbait after finding a real gap between what people clicked on and what they said they actually wanted to see. Clickbait got lots of clicks, which caused the ranking system to show it to more people. But when Facebook surveyed users, 80% said they preferred headlines that actually helped them decide whether to read the story.

That’s a really important distinction.

Engagement and satisfaction aren’t the same thing.

Facebook kept working on the problem. In 2016, it built systems to identify the linguistic patterns of clickbait. In 2017 it refined them again.

But what I find interesting looking back isn’t just that platforms eventually figured out how to detect “YOU WON’T BELIEVE WHAT HAPPENED NEXT.”

It’s what happened next.

Publishers learned. Creators learned. The content got better. The optimization got more sophisticated.

Upworthy itself didn’t survive the shift — once the curiosity-gap headline stopped working, the whole model underneath it collapsed.

Which is why I wonder if we’re making the same mistake with AI slop.

We’re focused on the version that’s easiest to see.

Slop is the easy problem

Right now, a lot of AI-generated content is obvious.

Six fingers. Weird faces. Strange phrasing. That oddly familiar cadence in a LinkedIn post.

Platforms are already building tools to catch it, and I’ve been in the middle of that conversation. The day after Substack rolled out its transparency feature, built with Pangram — letting readers scan longer posts, notes, and replies to estimate how much was AI-assisted — I posted a note pushing back on part of it. I’m not wild about how detection tools flatten the nuance of ethical AI use, but I’m a firm believer in transparency.

Here’s what that looks like in practice: one of my own posts scans as 100% AI-assisted under Pangram. I left the label up. I use AI to brainstorm, draft, edit, and summarize — but the ideas, the analysis, and the connections are mine, and every piece is verified and written in my voice. A 100% score doesn’t mean the words aren’t mine. It means I used the tool the way I said I would.

LinkedIn followed with something blunter: a “Seems like AI slop” button that members can click on any post, which trains the platform’s own detection models.

People mock the obvious stuff. We scroll past it. That’s good.

But what happens when the content isn’t bad anymore?

When it’s genuinely funny? Or useful? Or so precisely matched to your interests that you don’t particularly care how it was made?

That’s AI candy. And no detection tool built for spotting six-fingered hands is built for spotting that.

There’s already some evidence we’re drawing different lines depending on what the content is. A Meltwater/YouGov survey of nearly 10,000 consumers across seven countries earlier this year found that 53% considered generative AI acceptable in entertainment content.

For news reporting? 21%.

Political advertising? 18%.

And 86% said AI-generated content should be disclosed.

I think that’s a pretty reasonable instinct.

I’m not arguing that AI-generated entertainment is bad. There is some genuinely delightful stuff being made with AI, and there will be much more of it.

I actually have no problem with an entertaining AI influencer if I know it’s AI.

The questions start when I don’t.

The Macedonian teenagers, except personalized

The week after the 2016 election was a blur inside Facebook. Stories surfaced showing fake news outpacing real journalism in engagement — fabricated headlines like “Pope Endorses Trump,” some of them coming out of Veles, Macedonia, where teenagers had discovered they could make real money flooding our system with political clickbait. Samanth Subramanian’s WIRED investigation later laid out how: more than 100 pro-Trump sites registered to that one town, in a country where the average monthly salary was $371.

The detail that has always stuck with me wasn’t the money those teenagers made. It was the amorality of it. They didn’t particularly care who won. They figured out what Americans would click on, produced more of it, and made money.

Now take that same incentive structure and give it generative AI. My friend Renee DiResta has been talking about this a lot.

Imagine an AI influencer who looks like someone you might know. Talks like you. Likes the same sports team like you do. Maybe likes some of the same music, follows some of the same creators, has the same sense of humor.

You encounter them over and over again.

You start to trust them.

And you have no idea that the person operating that account is sitting in another country. Maybe they’re trying to influence American politics. Maybe they’re pushing extremist content. Maybe they don’t care about any of it and are simply discovering that outrage pays better than recipes.

And unlike the Macedonian teenagers, they can test almost endlessly — which face works, which political message, which version of the story keeps you watching for another 20 seconds.

That’s the part I don’t think we’re talking about enough.

The candy factory is already being built

And this isn’t speculation about some distant AI future.

Zuckerberg has been remarkably explicit about where Meta thinks this is going.

In October 2024, he told investors that social media had gone through two eras — first content from friends and family, then creator content — and that AI was about to add “yet another huge corpus of content on top of those.”

Then this summer he went further.

Meta’s new systems, he said, will be able to understand more deeply what people are interested in and what their goals are when they’re using the apps. They’re getting even more data about us with Muse, Meta’s newly launched set of AI models. (Full disclosure: my AI assistant is also built on Muse — an agent that helps me research and draft, not Meta’s image model. Welcome to the era of confusing AI branding.)

The result? That “whole new and nearly infinite universe of personalized content.”

Put those pieces together. Generative AI is dramatically lowering the cost of making content. Recommendation AI is getting dramatically better at understanding what each of us finds compelling. And the business model still rewards attention.

That’s the combination I’m watching. Not AI slop. AI candy.

Nobody’s decided who’s responsible for the candy

This is where we get into the impossible tradeoffs, because I don’t think the answer is “ban AI content.” Far from it. But there are questions we’re going to have to answer.

Should platforms monetize AI-generated influencers the same way they monetize human ones? What should an AI influencer have to disclose about who or what is behind it? Do media companies create their own AI personalities? What happens when an AI personality moves from entertainment into news or politics? Will TikTok, YouTube and X follow Meta into increasingly AI-generated feeds? And eventually: do OpenAI, Anthropic and other AI companies decide that the feed itself is the product?

I don’t know the answers to all of those. We’re still early enough to have the argument before the defaults get set for us.

We spent a lot of time after 2016 asking how we missed the Macedonian teenagers.

I think some of today’s AI influencers will turn out to be the Macedonian teenagers of 2016.

Disrupting Politics: A Front Row Seat to the Collision of Technology and Democracy, coming out October 6, is now available for pre-order in paperback and e-book. An audio version is coming.