From 2 August 2026, Article 50 of the AI Act applies in the European Union — the part of the regulation that imposes transparency obligations on certain systems and certain content. Anyone who has a person interact directly with an AI system must tell them so, unless the artificial nature of the interaction is self-evident. Providers of generative systems must make certain synthetic outputs detectable through machine-readable markings. Deployers — meaning those who use those systems in a professional activity — must make deepfakes and generated or manipulated texts on matters of public interest recognisable when they are published without substantial human review or editorial oversight. For generative systems already placed on the market before 2 August, a transitional period runs until 2 December 2026, but it covers only the obligation to apply technical markings to outputs.
The regulation does not define "slop". It introduces no general quality standard. It does not require that every text touched by a model always display a label visible to the public. It also draws a distinction between a notice directed at whoever is interacting with a system, a machine-readable marking, and a disclosure perceptible to the public. It makes the artificial nature of something recognisable in defined cases. It does not certify truth, originality, or value.
Around the same time, LinkedIn announced a different system. Hari Srinivasan, the company's Chief Product Officer, unveiled new classifiers to identify posts deemed "AI slop" or generically low quality and limit their distribution beyond the author's immediate network. Some users can flag that a piece of content "looks like AI slop"; LinkedIn is also testing private warnings to authors when their posts are perceived as inauthentic or overly AI-dependent. The geographic scope and rollout share are not public.
For Srinivasan, using AI and producing slop are not the same thing, and the definition of slop itself shifts. LinkedIn claims that in initial tests its classifiers identified generic content with 94% accuracy. The figure is company-supplied and comes with no public audit. Separately, the company says it wants to reduce the distribution of content classified this way.
The two things can overlap, but they are not the same. The AI Act sets recognisability and disclosure obligations in defined cases. LinkedIn uses classifiers and user feedback to decide the reach of posts. What follows is a possibility — not yet documented by any specific case: a piece of content can comply with all applicable European obligations and still be judged generic by the platform, and therefore receive less distribution.
It is a power that is harder to describe cleanly, because it does not present itself as law and often disguises itself as feed hygiene. But it touches the work of everyone who writes, publishes, sells, and competes for attention online. The next time someone tells you that the law "labels AI content", ask them which law, which content, which label. Then ask LinkedIn by what criteria it decides to reduce the reach of a text that complies with all applicable obligations.
