LinkedIn just announced that over a million people have used its “Seems like AI slop” button since launching it in late July. The platform is seeing a 40% reduction in views on posts classified as AI-generated, which sounds like a win. But I think there’s more to unpack here about what this really means for how we build, moderate, and trust content online.
The context matters here. Pangram AI’s detection tool flagged 41% of LinkedIn’s longform posts as fully AI-generated before LinkedIn rolled out these changes. That’s a staggering number, and it prompted the company to act. The “slop” button, combined with improved classifiers and the removal of LinkedIn’s AI “enhance” feature, represents a meaningful shift in how the platform approaches generative content.
The Signal vs the Noise
What I find interesting isn’t just that the button exists, but that a million people used it in such a short timeframe. That’s a clear signal that users care about authenticity, or at least perceive AI-generated posts as low-quality. Whether that perception is always accurate is another question entirely.
The real challenge with content moderation at scale is that you need both automated systems and human feedback loops. LinkedIn’s approach of combining ML classifiers with user reporting creates a virtuous cycle. Users flag posts, the platform learns what constitutes “slop,” and the classifiers improve over time. This is how content moderation actually works at scale, and I respect that LinkedIn is being transparent about the numbers.
But here’s what concerns me: the reduction in views could create a chilling effect on legitimate uses of AI writing assistance. Someone using GPT to help draft a professional post about industry trends shouldn’t be penalized the same way someone mass-generating engagement bait gets penalized. The distinction between “AI-assisted” and “AI-generated” matters, and I’m not sure any classifier can reliably make that call consistently.
The Authenticity Problem
This whole situation points to a deeper problem in how we think about social media platforms as information sources. LinkedIn has always been a hybrid: part professional network, part personal brand showcase, part content distribution channel. When 41% of posts are flagged as potentially AI-generated, we’re not just seeing a moderation problem. We’re seeing a fundamental identity crisis.
Users are optimizing for reach and engagement, and AI tools make that easier. LinkedIn’s algorithm has historically rewarded certain types of content, so people game it. Add generative AI to the mix, and you get mass production of “authentic-sounding” professional hot takes. It’s efficient, scalable, and increasingly hollow.
The company’s stated philosophy of “approaching this assuming good intent” is admirable, but I wonder how sustainable it is. If 40% less visibility on AI content is the result of this gentle approach, what happens when people get better at hiding AI generation? We’re in an arms race where detection gets better, obfuscation gets better, and both sides keep iterating.
What This Means for Developers
If you’re building tools that interact with LinkedIn’s API or content ecosystem, you need to think about how your application handles this shift. If users are increasingly sensitive to AI-generated content, and LinkedIn is actively suppressing it, then any tool that helps you post at scale without clear human involvement becomes riskier.
More broadly, this is a case study in platform policy evolution. Every social network will face similar questions about AI-generated content. Twitter, Reddit, Threads, and others are all wrestling with the same problems. The winners will be platforms that find a way to distinguish between legitimate AI assistance and low-effort spam generation.
There’s also an interesting developer angle around detection itself. If you’re working on classifier models or content moderation systems, LinkedIn’s success here (40% reduction in AI slop views) suggests that combining multiple signals works better than any single approach. User feedback + ML classifiers + feature removal creates redundancy in a good way.
The longer-term question is whether this approach scales. LinkedIn has the advantage of a professional network where people have reputation and identity tied to their accounts. That makes moderation harder to game. But it also makes people care more about their digital footprint, which is why they’re clicking the slop button in the first place.
What happens when the cost of creating AI content approaches zero, but the cost of getting caught distributing it remains high?