LinkedIn's New "Seems Like AI Slop" Button Is a Vote, Not a Detector

Ever noticed LinkedIn's new "seems like AI slop" button? It lets any member flag something, and those flags feed into the feed ranking system. That doesn't mean LinkedIn thinks the button is spotting AI. Mixing a stranger's instinct with a measured detector score ends up messing with the algorithm and can throw off user experience. Independent detection research puts a huge gap between a trained classifier and a gut feeling. Best-in-class commercial tools stay almost zero on false positives with real writing. Guesswork alone can't match those guardrails. So this button isn't meant to replace detectors; it's a way to see how common slop actually is.

What LinkedIn actually shipped

404 Media reported on July 30, 2026 that LinkedIn added a new option to its post menu titled "Seems like AI slop," letting users flag content they suspect was machine-written. LinkedIn's chief product officer, Hari Srinivasan, confirmed the feature himself, posting on the platform: "We are ramping up a series of new and improved classifiers that identify if a post is AI-slop or generally low-quality content. This will reduce the amount of AI slop you might see in suggested content and content from outside your network."

The mechanism works straightforwardly. It is not a detector that runs on the post itself. Rather, it uses a crowdsourced report. That report enters a signal for LinkedIn's internal ranking classifiers. The signal functions the same as a spam report or an "I don't want to see this" click. LinkedIn quietly retired its "enhance with AI" post-drafting tool around this time. The tool was replaced with a lighter proofreading option that does not rewrite content wholesale. LinkedIn's content policy was updated to explicitly favor professional conversations over low-quality, automated or generic content.

Pangram, a commercial AI-detection vendor, estimated that roughly 41 percent of LinkedIn's long-form posts are likely AI-generated. In short-form posts the figure drops to about 30 percent. The initial report came from 404 Media. LinkedIn's leadership has made slop a top priority. According to Srinivasan, the platform is treating slop as a top priority.

A flag button is not a detector, and the difference isn't academic

A detection tool gives each passage a score derived from measurable statistical properties such as perplexity, burstiness and the structural patterns typical of model-generated text. That score is validated by feeding the detector thousands of examples whose origin, human or AI, is already known and recording how often it misclassifies in either direction. When a user clicks "seems like AI slop," no statistical test is run. The judgment relies on a vague vibe built from a few viral instances of poor corporate LinkedIn-style writing. That vibe carries no stated accuracy, no false-positive rate, no calibration and provides nothing that can be audited.

We know what a real detector's error rates look like because researchers have actually measured them. A 2026 working paper from the University of Chicago's Becker Friedman Institute, "Artificial Writing and Automated Detection," built a 1,992-passage test corpus across six genres, including résumés and reviews, the exact kind of professional writing that fills LinkedIn feeds, and matched human-written text against passages generated by four frontier models (GPT-4.1, Claude Opus 4, Claude Sonnet 4, Gemini 2.0 Flash). The researchers tested three commercial detectors (Pangram, Originality.ai, GPTZero) and one open-source baseline (RoBERTa) for exactly the two errors that matter in a policy setting: false positives (flagging real human writing as AI) and false negatives (missing real AI writing).

The data show a large disparity among detectors once you set aside the difference between a detector and a human guess. Pangram reports "essentially zero" false-positive and false-negative rates on medium-to-long passages, and the rates rise only slightly on short text. The open-source baseline, meanwhile, is "unsuitable for high-stakes applications," misclassifying most human text with false positive rates ranging from about 30% to 78% across scenarios. A clear spread appears among the tools that attempt to measure something quantifiable. Even if many users click a button based on first impressions, they still fall short of the performance level set by the weakest detector in the study.

What the button is actually good for, and what it isn't

The flag button addresses a separate concern than a detector does. LinkedIn does not seek forensic certainty over a single post. Since it updates its ranking classifier using an aggregate signal from millions of posts, it is not dependent on one piece of evidence. The noisy crowd input serves as a cheap yet valid data source for that task. Srinivasan's framing affirms that the objective is to hone models and feeds, not to resolve ownership on a per-post basis. Using the flag button for model tuning and feed adjustment is warranted.

Where it goes wrong is if it gets treated as evidence about any one specific post or person, the way a detector score sometimes gets treated in a classroom or a newsroom. If a colleague's post gets mass-flagged, that tells you something about how the post reads, not what actually produced it. We've written before about the AI "tells" people look for in writing, the overly tidy structure, the relentless "Moreover," the em-dash-heavy rhythm, and those tells are real patterns worth noticing. They're also patterns that plenty of careful human writers produce naturally, especially non-native English speakers and people trained in formal business writing. A crowd of strangers pattern-matching on the same handful of cues, with zero calibration against real outcomes, is exactly the kind of confirmation-prone judgment that a properly built detector is designed to correct for, not replicate.

The honest reading of "41 percent of posts are AI slop"

It's worth pausing on that Pangram-sourced figure too, because a single headline number like "41% of long posts" can do the same kind of flattening a detector's own aggregate score does. We've covered why a single percentage hides more than it reveals: a post that's 40 percent AI by one measure might mean a fully generated draft with light human edits, or a human draft polished by an AI grammar pass, and those are very different situations for a reader trying to judge trustworthiness. The Pangram estimate is a useful directional signal about how saturated a platform has become. It isn't a verdict on any individual post, and neither is a stranger's slop-button tap.

Flagging content that appears to be AI-driven is far from futile; it still has a role. The choice of tool must reflect the gravity of the outcome it feeds. A social media platform can rely on cheap, noisy signals from a broad audience since an error is marginal. When the stakes involve employment offers, scholarly integrity, or public attribution, the ramifications are substantial for a real person. That's the same gap we've written about before: even a detector that has been carefully calibrated begins the investigative process, it should not replace human assessment. A pile of button clicks sits well below the threshold of that already-limited standard.

FAQ

Does LinkedIn's slop button use real AI detection technology? The flag doesn't function as a trained classifier scoring the text of a post. It is a signal that a user submits, and this information is used by LinkedIn's ranking and moderation systems. Its purpose mirrors that of a spam report. LinkedIn has said it is developing improved classifiers that automatically catch slop. So the button simply serves as a crowd signal, not a result from a detector.

Is the "41% of LinkedIn posts are AI-generated" figure reliable? A vendor estimate from the detection firm Pangram appeared under a 404 Media report. Afterward, the story was picked up by outlets such as The Independent. It should be noted that this is not an independently peer-reviewed study. The estimate suggests the platform may face a real saturation problem, but it does not provide a precise, audited statistic.

Can I trust a coworker's judgment that a post "seems AI-generated"? Treat any single opinion as a data point, not a verdict. A human pattern-matching AI captures real cases when it encounters such signals. Yet the same AI also flags many human writers who adopt a formal, tidy, or transition-heavy style. A measured detector score, read as one signal among several, proves more reliable than an unstructured hunch. Neither the detector score nor the unstructured hunch should drive a high-stakes decision about a real person.

Why does detector choice matter so much if they're all "AI detectors"? Because the false-positive and false-negative rates between tools vary enormously, not narrowly. The Chicago Booth working paper found error rates ranging from "essentially zero" to 30-78% depending on the tool, all claiming to do the same job. A detector's actual, measured accuracy matters more than the fact that it's labeled a detector at all.

Curious whether one specific piece of text was actually written by AI? Skip the vibe check. Run it through our detector and see the sentence-level breakdown behind the score.