AI Slop, Watermarking, and the Rise of Commodity Content

Most comments rolling into my LinkedIn notifications these days follow a familiar pattern. Slop. Naturally. It has become the default response to nearly everything I publish there.

The biological metaphor fits better than it should. A virus carries no metabolism of its own; it needs a host to reproduce. Slop comments do the same thing, echoing my own words back at me with nothing added, no insight, no friction, no value. My enthusiasm for spending time on LinkedIn has eroded to roughly the level of enthusiasm I hold for a dental appointment. The platform made it this way when it chose to enable and encourage AI-generated participation.

LinkedIn, to its credit, sees the infection spreading. Over the past ten weeks, the platform has rolled out a defensive response:

Antibodies perform three jobs: they identify a target, neutralize it, and retain a memory of it so the next encounter is faster. Anti-slop systems follow the same playbook. They flag low-value content, throttle its distribution, and feed every detection back into the model for future refinement. LinkedIn is effectively building Slop Antibodies, and it is far from alone. Every major content platform is engineering its own immune response:

Substack now applies Pangram across the entire site to flag AI-written material. YouTube has moved to demonetize content deemed “repetitive, low-effort, emotionally manipulative.” In January 2026, it terminated 11 channels and wiped six more, erasing roughly 4.7 billion lifetime views, 35 million subscribers, and about $9.8 million in annual revenue. Reddit chose the anti-manipulation route over labeling. As of July 2026, its AI-based detection targets “manipulated and spammy content,” reporting 23 million spam views blocked and roughly 2 million inauthentic votes revoked each day. Pinterest combines AI detection with labeling and gives users a feed control to dial down “AI-modified” content in specific categories. TikTok requires AI labels, embeds invisible metadata watermarks, shipped a “limit AI content” feed toggle in November 2025, and began testing detection aimed at accounts dedicated to AI spam in July 2026. Meta has applied “AI info” labels across Facebook, Instagram, and Threads since 2024, extending them to ads in June 2026, though no user-side filter exists.

Spotify attacked the supply side directly: over 75 million spammy tracks removed, an impersonation policy, spam filters, and DDEX-based AI disclosure in credits.

Cheap synthetic content floods the platforms, and the immune systems react with labels, reporting flows, filters, demonetization, spam detection, and supply-side removals. LinkedIn claims its defense system operates at 94% accuracy. That sounds impressive until you compare it to Gmail’s spam filters, which are roughly 60 times more precise. At LinkedIn’s rate, one in 17 spam items would surface as a false positive.

This shift reshapes demand in fundamental ways. Production costs collapse toward zero, which pushes the pressure onto distribution as it moves under machine control. The scarce asset becomes selection. Content production efficiencies mean nothing if the content never reaches the right audience.

Watermark panic

On August 11, Anthropic announced machine-readable watermarks on Claude text and file output at the model level, applied worldwide. The marks persist across the API, Claude, Claude Code, and cloud access through AWS, Google Cloud, and Microsoft Foundry.

Anthropic did not act out of altruism. The EU AI Act’s Article 50 obligates providers of generative systems to mark their output in a machine-readable format, with penalties up to €15 million or 3% of global turnover. Every lab serving the EU owes that duty, whether or not it also signed the Commission’s voluntary code on top.

Anthropic’s watermarking is the upstream version of the same immune response content platforms deploy. Instead of asking LinkedIn, YouTube, or Reddit to infer whether a post was generated by AI after it enters the feed, Claude can make generated text easier to identify at the source.

I am not concerned about the impact of watermarking on marketing:

Watermarks only prove that an AI model touched text, not to what extent. They are not an indicator of quality.

Detectors need a minimum amount of text to function. Published benchmarks land at roughly 100 tokens at best, and SynthID’s own evaluation truncates everything to 200. A LinkedIn comment is 20 to 50 tokens. The slop I opened this piece with sits below the detection floor.

Editing, paraphrasing, translating, combining the response with other text, or chaining models can weaken or remove the mark.

If Anthropic makes watermark detection publicly available, it will trigger a cat-and-mouse game in which actors reverse-engineer ways to break the watermark, and Anthropic must figure out how to harden it.

A paraphrasing attack presented at ICML 2025 achieved near 100% success against seven recent watermarking methods at a cost of $0.88 per million tokens.

The escape hatch is open weights. Watermarking is applied by the sampling pipeline at inference, so if you run the model yourself, there is nothing to strip. If marked output ever gets penalized, slop moves to the models nobody marks.

Fun fact: Gemini has been applying SynthID watermarks since August 2023, and there was no outcry. The web’s reaction to Anthropic’s announcement says a lot about Google’s AI position.

At the heart of this fear of watermarking is my key argument: We should not confuse “AI-generated” with “bad.”

Slop existed before AI. Poor or machine-automated work, legal and finance speak, press releases, and every SEO article that opened with “in today’s digital landscape” all qualify. It is ultimately about quality. Where the fear is justified is that quality has become a blurry but critical filter for distribution.

Distribution bottleneck

The advice to avoid generic content is not new, just widely ignored. But AI raises the viral load faster than the filter can clear it.

Content production is no longer the constraint. Permission to distribute is.

What we learned over the last 24 months is that the production gains from AI do not offset the losses in distribution. You could argue that AI Overviews reduce clicks by 50% on average, but you can grow content output 2x to compensate. But that output growth comes at the expense of quality, which can ultimately hurt your overall distribution. An article or post that smells like AI can cost you fragile trust with readers to an irreparable degree.

The census above was platforms protecting their own feeds. The layer that decides what AI answers cite is doing the same thing, one level up.

Google shipped a spam update in June aimed at scaled content abuse. Wikipedia went furthest: speedy deletion for suspected LLM-generated articles in August 2025, then an outright ban on using LLMs to write or rewrite article content in March 2026.

Google demotes, Reddit detects, Wikipedia bans.

Shouldn’t AI Slop hurt the distribution of content platforms themselves? After all, LinkedIn, Reddit, and YouTube are the most-cited domains and bigger slop slingers.

Why are they not penalized by search engines and LLMs?

They are, but more targeted. Reddit’s machine-translated pages collapsed from 6.14% of ChatGPT’s Reddit citations in April 2026 to 0.30% by early June, while Reddit rose in aggregate over the same period.

We do not know what LLMs filter out for training data versus live retrieval.

Engagement may act as a rough filter. Semrush’s study of 89,000 cited LinkedIn URLs found the cited ones carry at least decent engagement.

Most cited posts have moderate engagement (15-25 reactions), while about 75% of cited authors post frequently (5+ posts in four weeks) and nearly half have over 2,000 followers.

So it seems that at least some sort of slop filtering of content platforms is happening. But what is the cost to distribution for companies? Several studies looked into this.

A Copenhagen Business School study published in Electronic Markets ran two experiments (n=325, n=371) on Instagram content labeled human-created, AI-enhanced, or AI-generated. It found that labeling as AI-generated or AI-enhanced reduced both affective and behavioral engagement by about half, with the effect strongest on emotional content and weakest on rational or informational content.

TikTok field data from 1 million posts shows AI disclosure leads to roughly 7% less engagement because people infer lower effort. Pretty tame in my mind.

Pangram scanned over 1 million posts between April and June 2026 and found that 41% of LinkedIn long-form posts and 23% of comments were fully AI-generated, the highest of any platform.

Run the numbers yourself. LinkedIn says it catches 94% of slop and caps flagged posts at your immediate network. In my analytics, 71% of impressions come from beyond that network.

Multiply the two: an account posting nothing but AI should expect to lose roughly two-thirds of its LinkedIn reach. Post AI a third of the time and it lands closer to 20%. Add 7% on short-form video, and 40% to 95% in SEO if you scaled a content operation.

That LinkedIn figure is derived from two published numbers and my own baseline. LinkedIn published a catch rate and nothing at all about false positives.

Commodity content

At the beginning of the article, I mentioned that LinkedIn slop comments regurgitate my original post and provide zero value. That might remind you of the concept of information gain and commodity content, and it should. In 2021, I wrote:

“Second, if the key feature is easy to replicate, you have a problem. You have a commodity; you’re one choice of many. Commodities compete heavily on price and cost. As a business, you have to weigh the cost it takes to create the product against the returns it brings. Are you in a stronger position when you compete with minimal advantage for a small share of ad views?”

Fast forward to today. Danny Sullivan shared an important slide at the Search Console Live Toronto conference about commodity versus non-commodity content.

In Sullivan’s words, non-commodity content is:

Unique: Brings a viewpoint, information, or has content that others lack or cannot easily replicate.

Specific: Talks about a specific instance, situation, or thing, not general rules, steps, or generic information.

Authentic: Demonstrates first-hand knowledge or experience.

It hits the whole slop discussion on the head. It is the universal anti-slop recipe. Look at how similar LinkedIn’s VP Laura Lorenzetti’s frame of AI Slop is to Google’s:

“This includes technology systems built in partnership with our editorial team that have been trained to recognize signals of AI slop and learn over time by identifying content that adds perspective, context, or expertise and content that feels generic or repetitive, even if it appears polished on the surface.”

Or what Reddit CEO Steve Huffman recently said on the company’s Q2 earnings call:

“As AI makes information more abundant, the challenge is no longer finding content; it’s finding context, personal opinion, and first-hand accounts. Everything online feels flat, polished, generated, or sponsored, so consumers are overwhelmed and increasingly skeptical.”

Three organizations arriving at the same conclusion: generic content has no value. Is that new? No.

What is new is the binary nature and cost-effectiveness. Generic content could gain some residual traffic until a few years ago.

Now, Google will not even index it, and LinkedIn downranks it. Panda ran on proxies: links, clicks, dwell time, pogo-sticking, because judging whether a document contained a good idea was economically impossible when it launched in 2011. Now, an LLM can assess this quickly and at low cost.

A COLM 2026 paper from Maryland and Google DeepMind ran 61,608 stories through a classifier that was deliberately denied every style signal and told to work from structure alone. It kept over 97% of the accuracy of models that were allowed to see word choice and sentence rhythm. The finding: slop has a shape pattern.

Now scroll back up and look at the shape of my LinkedIn comments from the beginning. Spotting the uniform shape?

There are even more similarities between the COLM study and slop: AI states the moral outright in 77% of stories versus 52% for humans, and 79% of AI stories contain zero subplots versus 57% for humans. Over-explaining and single-track tidiness is exactly the LinkedIn comment pathology.

So the antidote is owning something a model cannot produce on demand:

Proprietary information: Your own data, your own experiments, your own customer conversations. If a model can generate it, it is a commodity by definition.

Attributable identity: A named author with a track record, and on LinkedIn, a verified one. Lorenzetti’s post buried the tell: you can now filter the feed for LinkedIn’s 100 million+ verified members. Verification is an antibody.

An owned channel: Email, community, direct relationships. Somewhere the filter is not standing between you and the reader.

The test for all three is the same: would it be expensive for someone else to fake?

Wish for the filters to work. Every throttled slop comment is oxygen back for someone who actually had something to say. Just make sure that when the feed finally clears, you are the source the machine decided to keep.