Zenfeed
Every so often someone asks me, half-joking, "what percentage of this app is bots now?" I never have a real answer, and neither does anyone else, because nobody is actually counting. The number that would answer it doesn't exist anywhere I could find — not because it's a secret, but because it's genuinely hard to measure. So I decided to do the small, honest version myself: count my own feed by hand, for a week, and see what came out the other end.
I want to be upfront about what this is and isn't before a single number shows up. It isn't a study. It's one person, one feed, one week, using a checklist of tells rather than a real detector. If you're after a citable statistic, this isn't it, and nothing I could run on my own laptop would give you one. What it is: an honest look at what counting actually turns up, done once, by hand, with the method shown rather than hidden.
How do you even try to count AI-written posts in a feed?
By hand, using the same tells I wrote up in how to tell if a LinkedIn post was written by AI: a reversal-shaped hook, invisible-bullet-point paragraphs, a sentence rhythm that never stumbles, and a closing question aimed at nobody in particular.
Each day, I scrolled my usual LinkedIn feed once, the way I normally would, and noted which posts showed two or more of those tells together. I didn't stop to judge in the moment, because that changes how you scroll — I just marked a post and kept going, then reviewed the marks at the end of each day. By the end of the week I had a small, informal tally: how many posts I'd scrolled past in total, and how many had tripped two or more tells. That ratio is the "number" below, and I want to be precise about what it is: a hand count of one feed, against a checklist that catches a pattern, not a verdict.
- A hook built on a clean reversal — "I used to think X. Then Y happened."
- Paragraphs that read like bullet points even without the bullets.
- A sentence rhythm that never breaks stride — no fragments, no false starts.
- A closing question that isn't really asking anyone anything.
So what's the actual number?
About one in five of the posts I scrolled past that week showed two or more of those tells together — not a measured rate of anything, just one feed, counted by hand, for seven days. Take the method more seriously than the fraction.
I'd expected it to feel higher than that in the moment, and lower once I actually counted. It landed closer to my gut feeling than I wanted it to, which is its own kind of answer: the sense that a feed has quietly filled up with lookalike posts isn't always an exaggeration. Some days the count was closer to one in eight. One day, after a popular post about "lessons from a layoff" clearly spawned a wave of similar-shaped responses, it was closer to one in three. Averaging those days into a single week-long number flattens something that mattered in the moment — the volume comes in bursts, not evenly.
Does the exact percentage even matter?
Not really, and that's the point this whole exercise kept making. No honest hand count of one feed, done once, can tell you the true rate across LinkedIn, and a slightly different week would have given me a different number without anything about the platform actually changing.
What a week of counting is actually good for isn't the fraction — it's noticing that the fraction is high enough to bother counting at all. I didn't need a precise percentage to notice that on the loud day, nearly a third of what I scrolled past felt interchangeable. The number that matters isn't "22%" or "31% on a bad day." It's that either number is large enough to start crowding out the posts from people I actually follow, which is the actual complaint. None of this is a claim that any single one of those posts was bad, written badly, or not worth reading on its own. The problem shows up at volume, not in any individual post.
What would it actually take to measure this properly?
A large, randomly sampled set of posts, a detector better than a human checklist, and a method published openly enough for someone else to check — on a platform that doesn't make that kind of bulk access easy to get. I don't have any of that, and I'd be skeptical of anyone who claimed a precise number without it.
What I have instead is the smaller, honest version: the method shown in full, the sample tiny, and the result labeled for exactly what it is rather than dressed up as something more rigorous.
That's also roughly the position Zenfeed's own AI detection is in, and it's worth saying plainly rather than implying something cleaner. Rest from AI works by scanning each post's text, plus image alt text, against a list of AI-related phrases and words — it doesn't run a model that judges writing style, and it doesn't claim to catch every AI-written post or only AI-written posts. It's a volume valve tuned by a phrase list, the same kind of imperfect, visible method as the hand count above, not a verdict machine.
A timer that hides AI-flagged posts for a set number of minutes, with a live countdown badge on the toolbar icon.
Block posts by your own word list, for the specific patterns a phrase list alone won't catch.
On LinkedIn and X, Rest from AI runs the way it's described above. On YouTube it does less on purpose — the same keyword-style AI sweep, but only on Home, Subscriptions, and search results, with no equivalent for a watch page or a channel page yet. None of it is a precision instrument, on any of the three, and I'd rather say that outright than let a countdown timer imply more confidence than the method actually has.