LinkedIn exec details how it detects and demotes AI slop

LinkedIn now has, by its own admission, a standing internal effort dedicated to finding and suppressing "AI slop," the platform's term for low-effort, machine-written filler posts. Tim Junka, the company's Vice President of Engineering, laid out the mechanics in a LinkedIn post that Social Media Today reports on: human-trained classifiers, feedback loops from member reports, and AI agents built specifically to spot other AI's output. The admission matters more than the detail. LinkedIn is telling its own user base, in public, that it considers synthetic content a threat to the product, not a curiosity.

That is a notable reversal of tone. Through 2023 and 2024, LinkedIn's own product moves (writing suggestions, post rewrites, comment drafts) pushed members toward AI-assisted content. Now an engineering VP is describing a detection system built to catch the downstream effect of exactly that behaviour. The platform encouraged the fire, and is now staffing a brigade.

The tell is in who gets trained, not what gets banned

Junka's account, per Social Media Today, centres on classifiers trained against patterns: repetitive sentence structures, generic hook formulas, engagement-bait phrasing, the kind of content that reads like it was assembled to satisfy a prompt rather than convey a specific fact or argument. This is a pattern-detection problem, not a disclosure problem. LinkedIn is not primarily asking "was this written by AI." It is asking "does this look like the thousands of other posts generated to perform well rather than say something."

That distinction should worry anyone running a company page or executive profile on a templated content calendar. The industrial-scale approach to LinkedIn content that agencies and in-house teams have defaulted to over the past two years (take an announcement, run it through a formulaic hook-plus-bullets structure, repeat weekly) produces exactly the statistical signature a classifier is built to catch. It was never really about whether a human or a model typed the words. It was about whether the post resembled the thousands of others optimised for the same mechanical cues. LinkedIn has now confirmed it is actively scoring for that resemblance.

What gets demoted when a classifier can't read intent

The mechanism Junka describes cannot distinguish a thoughtful executive using AI to tighten a draft from a content mill churning out ten posts an hour. It works on surface pattern, not provenance. That is a problem for every brand that treats LinkedIn output as a volume exercise: the quarterly results post with the same three-bullet structure as last quarter's, the "excited to announce" format repeated across fifty company pages, the leadership-insight post that reads like forty other leadership-insight posts because it was generated from the same prompt template.

For financial services and multilateral institutions, this changes the calculus on in-house content operations. A communications team at a development bank or an industrial group producing weekly thought-leadership posts at scale, using AI tools to hit a cadence target, is now competing against a system explicitly built to flag cadence-driven, template-shaped content. The brands that will keep their reach are the ones whose posts carry specific, checkable detail: a named project, a cited figure, an argument that could only have come from someone who actually did the work. Generic authority-building prose, however polished, is now the thing the algorithm is tuned to find.

There's a second-order effect worth flagging for executive comms teams specifically. Ghostwritten executive posts have become standard practice at large B2B organisations: a comms team drafts, an AI tool polishes, an executive's name goes on top. If LinkedIn's detection runs on structural pattern rather than authorship, the risk is not that ghostwriting gets caught. It's that templated ghostwriting, the kind run off a standard outline across a dozen executives at the same company, starts looking indistinguishable from slop at the pattern level, even when every fact in it is true and every word was reviewed by a human.

Junka's post is also a signal about where LinkedIn wants its feed to go next: toward content that rewards specificity and penalises formula, regardless of whether a model or a person produced the formula. That is good news for the executive willing to write (or dictate) something with an actual point of view, citing a real number, naming a real client, making a real argument. It is bad news for the comms operation treating LinkedIn as a cadence problem to be solved with templates and tooling. LinkedIn has just told you, through its own VP of Engineering, which of those two postures it is now built to detect.