LinkedIn says its filter for generic AI-sounding posts is 94% accurate, but outsiders have no way to check that number.

This explores how much to trust a platform's detector when it marks posts as AI-generated or low-effort. The short answer is that nobody outside the platform knows. LinkedIn says its filter for generic content is 94% accurate, but the figure is self-reported. It comes from unspecified testing, with no false-positive rate and no description of what was tested How often does LinkedIn wrongly flag legitimate posts?. It isn't even clear whether '94%' means that most flagged posts really are generic, or that most generic posts get caught. Those are very different promises to a human writer whose post gets limited Does LinkedIn's 94% accuracy apply to human posts wrongly limited?.

The missing false-positive rate matters more than it seems, because a single accuracy number hides how rare the target is. A measurement of 51 subreddits found machine-generated text only marginally present overall. It peaked around 9% in a few technical and support communities and came mostly from a small number of users How much machine-generated text actually appears on Reddit?. When most posts are human-written, even a small error rate on human posts can produce as many wrong flags as right ones. That Reddit figure is itself one detector's flagging rate, so measurements of AI content inherit the same problem: you are counting what a detector says, not what is true.

There is also evidence that detectors pick up on writing style rather than where the text came from. Fake-news detectors flag truthful LLM-written articles as fake and let human-written disinformation through. They learned that a certain polished style signals deception, and AI happens to write in that style Why do fake news detectors flag AI-generated truthful content?. A slop detector faces the same risk in reverse: a careful human who writes in a tidy, list-heavy, 'professional' voice may look exactly like what it was trained to catch.

Human judgment won't fix this. A review of 30 studies found people spot AI-made text, images and voices at roughly chance levels Can people reliably spot content made by AI?. When users accuse each other of posting slop on Hacker News and Reddit, the accusations don't track the prose features that actually separate AI from human text. They work as social gatekeeping, not detection Do AI slop accusations actually detect AI text?. Labels also carry weight on their own: readers trust unlabeled AI-assisted messages as much as human ones and turn skeptical only once origin is disclosed Do readers trust unlabeled AI-written messages as much as human ones?. A wrong label therefore does real damage.

The useful question to ask of any detector is 'compared to what ground truth?' The same gap shows up in a study of AI agents gaming their rewards, whose reported cheating rates can't be interpreted because the paper never says how cheating was labeled How were reward hacks labeled in this benchmark study?. An accuracy claim is only as good as the labels it was scored against. For the posts in question, those labels haven't been published.