LinkedIn has a slop button now

LinkedIn is rolling out a feedback option that lets you flag a post as "Seems like AI slop." The company has said publicly that AI slop is a top priority.

A professional network built a button for reporting that a post appears to have been written by a machine. Not spam, not harassment. Slop. That is where we are, and I want to talk about it because the same platform is where roughly everyone I know goes to establish themselves as an authority on AI.

The numbers, with the caveat that matters

Two detection studies came out in July and they disagree.

Originality.ai sampled 5,000 public posts across nine topics and classified 81.2% as likely AI. Pangram analyzed more than a million posts and found over 40% of posts longer than 250 words were entirely machine-written.

Those are very different figures and the gap is methodology. Different detectors, different thresholds, different definitions of what counts. I'm citing both because citing whichever one is more alarming would be exactly the behavior I'm about to complain about.

Either number is remarkable. Between them the studies also report that LinkedIn accounts for something like 62% of all AI-generated posts across the platforms sampled, making it the most synthetic place measured, and that the machine-written posts draw roughly 45% less engagement. An enormous volume of text is being generated that nobody reads.

Which raises the obvious question about me, so let me answer it before going any further. Do I use AI to help produce what I write? Of course I do. Most of what I've published this year passed through a model at some point, usually for a first pass or a structural edit. What I'd defend is everything that happens after that. A human reads every line before it goes out, checks the claims against sources, and throws away the parts that came back wrong, which on a bad draft is most of them. That step is not optional and it is not fast, and skipping it is the whole of what I'm complaining about.

Everybody became an expert in the same quarter

Scroll for four minutes and you'll find a dozen people confidently explaining agent architecture, context engineering, evaluation strategy, and the correct way to structure a multi-agent system.

Some of them are doing the work. Most of the confident ones were posting about something else entirely eighteen months ago, and the transition happened without any visible period of learning in between. There's no "here's what I got wrong while figuring this out" phase in the timeline. It goes straight from silence to authority.

Real expertise has a texture. It sounds like somebody who has been surprised recently.

My favorite version of the genre is the token spend post. Somebody screenshots a monthly API bill and presents the number as evidence of seriousness. Look how much inference I bought.

I'll put my own numbers up. A handful of subscriptions in the twenty to a hundred dollar range, run for a couple of years now, and I have never once hit a usage limit on any of them. I add metered API spend when a project genuinely needs it, always with a cap, and I have never been in a position where the work demanded a serious amount of inference. That covers everything on this site and everything I've built alongside it, during a stretch where I've been shipping harder than at any point in my career. If heavy spend were the entry fee, I'd have hit the wall a long time ago.

Spend measures nothing. It's an input. It tells you nothing about whether the output was correct, whether anyone used it, or whether the same result was available for a tenth of the cost with a smaller model and better prompts. A large bill is as easily evidence of a retry loop you haven't noticed. This is the same move as posting your hours worked, updated for a new unit, and it survives because a number feels like proof even when nobody has said what it's proof of.

The tell is specificity

There's a reliable way to separate the two groups and it takes about one sentence.

People doing the work describe specific failures. The tool call that silently returned a challenge page. The eval that looked great until you noticed the test set leaked. The thing that worked in December and broke when the model updated. They name the version, the error, and the amount of time they lost.

People performing the work describe categories. Paradigm shifts, inflection points, the death of some job title, three lessons from their AI transformation journey. Nothing in the post could be wrong, because nothing in it is specific enough to be checked.

The second kind gets more engagement, which is the entire problem.

The part that made me laugh

A few weeks ago I wrote a small script that scans my drafts for the phrasing patterns that read as machine-generated. Em dashes where a period belongs. The negative parallelism where you say a thing is not X, it's Y. Rule-of-three lists that were forced into three. A vocabulary I now find hard to unsee, all the delving and the underscoring and the tapestries.

I built it because my own drafts kept failing on those patterns and I was tired of catching them by eye.

Then I started noticing them in the feed, and the feed is saturated with them. The posts explaining how to think about AI are being written by AI, in the recognizable voice of a default assistant with the temperature turned down, by people whose entire pitch is that they understand this technology deeply.

I don't think most of them are being dishonest, exactly. I think they've automated a task they were never doing carefully in the first place, and the automation is more honest about that than they are.

Why it isn't only annoying

If this were purely an aesthetic complaint I wouldn't bother.

The cost falls on people trying to learn. Somebody two years into their career, trying to work out whether they should be running evals or building agents or worrying about their job, is reading a feed where confidence and correctness have come completely apart. The most certain-sounding posts are frequently the least grounded, because certainty is what the format rewards and grounding is expensive.

That was always somewhat true of professional social media. What's new is the volume. One person can now produce forty confident posts a week about a field they entered last spring, and the ranking system cannot tell the difference.

I am also on there

I post on LinkedIn. Something I wrote there last month travelled further than anything else I've published this year, and I'll be posting a link to this article on the same platform, which is a level of irony I'm choosing to accept.

I'm not above the thing I'm describing. My drafts fail my own script constantly. I've had to correct a bad statistic in a published post, and I only caught a second one because I opened the primary source out of paranoia rather than process. The difference I'm trying to hold onto is narrow: show the specifics, say when you were wrong, and don't post about things you found out about last Tuesday as though you've known them for years.

That's a low bar. Judging by the button LinkedIn just shipped, it's apparently high enough.