Original article by Kevin Indig
30-Second Rundown
Key learnings:
- AI made content cheap and abundant, so platforms are building systems that downrank or remove generic, repetitive output. More volume no longer guarantees more reach.
- A watermark shows that a model touched the material, not whether the result is useful. Distribution filters care more about perspective, context, expertise, and first-hand evidence.
- The defense is content expensive to fake: proprietary evidence, a named author, a clear point of view, and an audience you can reach without permission from a feed.
This week: Choose one recent post and mark every passage that merely repeats information already available elsewhere. Identify the three weakest sections, then replace each with your own data, a named customer example, a documented experiment, or a first-hand observation. Add a named author and a clear point of view so the finished post contains something a competitor could not reproduce with one prompt.
Coming 3 months: Build an automated pre-publication workflow that checks every draft for generic claims, missing evidence, weak author attribution, and passages that could appear on any competitor’s website. Route flagged sections back to the content owner for original data, experience, or opinion before publication, then use AI Search monitoring and rank tracking to measure whether stronger articles earn more visibility, citations, and reach.
Kevin Indig looked at the great synthetic-content flood and found the platforms growing an immune system in real time. AI made words, images, songs, and videos almost free to manufacture, so the scarce resource is no longer production. It is permission to reach another human before a machine decides your work smells like reheated paste.
Indig calls these defenses “slop antibodies”: systems that spot generic, low-value content, throttle it, and learn from each catch. The point is not that every AI-assisted sentence is diseased. The point is that repetition without perspective has become cheap enough to drown the town.
1. The Feeds Are Building Immune Systems
LinkedIn moved fast. In roughly ten weeks it deployed systems to identify slop and restrict flagged content beyond a poster’s immediate network, tested a user report button for AI slop, and killed its own rewriting tool in favor of a proofreader that fixes grammar without stealing the author’s voice.
Across the wider wasteland, platforms are choosing different weapons: labels, feed controls, demonetization, spam detection, vote reversal, and outright removal. When production becomes nearly free, distribution becomes the tollbooth, and the platforms now decide which truck gets turned back into the desert.
More output cannot reliably compensate for less reach, because rushing twice as much generic material through the pipe can damage trust and teach the filters exactly what to suppress. The content factory is no longer a growth engine if the loading dock has been sealed with concrete.
2. The Watermark Panic Is Aiming at the Wrong Corpse
Anthropic’s machine-readable watermarking made marketers nervous, and European law is pushing model providers toward marking generated output. But a watermark only says a model touched the material. It does not tell a reader whether the result contains evidence, judgment, experience, or one original thought.
Indig lists the weak joints: detectors need enough text, short LinkedIn comments may fall below the threshold, editing and translation can weaken the mark, and open-weight models can run through pipelines that add no mark at all. A cited 2025 attack reportedly broke seven watermarking methods almost every time for less than a dollar per million tokens, which means the cat already owns lock picks.
The durable distinction is not human versus machine; it is valuable versus disposable. Bad corporate prose, legal fog, press releases, and copy-paste search articles were slop long before a chatbot learned to put on trousers.
3. Generic Content Has a Shape, and the Machines Can Smell It
Indig connects the platform crackdown to commodity content: material that is easy to reproduce and therefore offers no reason to choose one source over another. Google’s Danny Sullivan describes the opposite as unique, specific, and authentic, while LinkedIn talks about perspective, context, and expertise. Three institutions walked into the same saloon and pointed at the same corpse.
A 2026 study from researchers at the University of Maryland and Google DeepMind tested whether a system could distinguish AI-written stories from human ones using only the structure of each story, not its vocabulary or writing style. Across 61,608 stories, that structure-only system retained more than 97% of the accuracy achieved by systems that could also analyze word choice and sentence rhythm.
AI stories often share a recognizable blueprint. The writing can look polished while still following a predictable, single-track shape.
The antidote is material expensive to fake: your own data, your own experiments, named authors with a visible record, and direct relationships through email or community. If another company can reproduce the article in one prompt without knowing your customers, your results, or your scars, you have manufactured a commodity and painted it beige.
Wish for the filters to improve. Every throttled paragraph of polished nothing gives oxygen back to the person who actually investigated, tested, interviewed, or noticed something.
Use the machine if it helps, but bring evidence, identity, and a point of view the machine cannot manufacture on demand. When the feed finally clears, the survivors will not be the loudest factories. They will be the sources worth keeping.
