Generative artificial intelligence has made producing content extraordinarily cheap.

A script, a LinkedIn post, an article, a synthetic voice, or even a complete video can be generated in minutes.

But that abundance is creating its own problem:

when producing content costs almost nothing, the Internet can also fill up with content worth almost nothing.

YouTube, LinkedIn, and Substack are beginning to react.

All three platforms have introduced mechanisms to identify, label, or reduce the presence of certain AI-generated content. This is not a general ban on AI. In fact, their own policies make it fairly clear that using AI as a tool can be perfectly valid.

The important change seems to be somewhere else:

separating AI-assisted creation from industrialized, repetitive, deceptive content or content with no meaningful human contribution.

A recent article by Amy Zwagerman in Inc., also republished and discussed by other outlets, captures the broader signal well: platforms are already building their own infrastructure to decide how to tell the public that content was created or altered with AI.

And that changes the rules of the game for creators, companies, and media organizations.

YouTube wants to know when what we see looks real but is not

Since 2024, YouTube has asked creators to disclose when they use AI to generate or significantly alter realistic content.

In May 2026 it took another step.

The platform moved labels for photorealistic AI-generated or AI-altered content into much more visible positions: below the player on long-form videos and directly over the image in Shorts.

YouTube also began rolling out internal signals capable of detecting certain uses of AI and automatically applying a label when the creator has not made the corresponding disclosure.

The policy focuses especially on content that:

  • makes it appear that a real person said or did something that never happened;
  • alters footage of a real event or place;
  • generates a realistic scene of an event that never occurred.

There is a fundamental distinction here.

Using AI does not automatically mean a video must be labeled.

YouTube’s own documentation lists several uses of AI that it considers production assistance and that, by themselves, do not require disclosure: generating or improving a script, developing ideas, creating a title, thumbnail, or infographic, producing captions, enhancing audio, or even cloning your own voice for narration and dubbing.

That breaks down an overly simple interpretation of these policies.

YouTube is not saying:

“Do not use AI.”

It is saying something more like:

“If you use AI to present something as real that did not happen, we want the viewer to know.”

There is another important nuance: YouTube explicitly says that the mere presence of an AI label does not change how a video is recommended or, by itself, affect its eligibility for monetization.

The problem, then, is not simply using artificial intelligence.

The problem starts when AI changes the trust relationship between creator and audience.

LinkedIn is going directly after “AI slop”

LinkedIn is facing a different problem.

On a network built around people sharing professional experience, the massive arrival of automatically generated text threatens to turn feeds into an endless collection of perfectly written and perfectly forgettable posts.

LinkedIn already uses an explicit phrase to describe that phenomenon:

AI slop.

The company defines it as content likely generated with AI that requires little effort, may look polished on the surface, but lacks a clear point of view, first-hand experience, or substance.

It can be generic, repetitive, recycled, or designed primarily to capture attention.

What is interesting is that LinkedIn is not defining the problem only by the tool used.

Its documentation states the principle fairly directly:

the focus is not on how the content was created, but on whether it provides value.

AI-assisted content remains welcome when it reflects the perspective, experience, or knowledge of a real person. Generic or repetitive content, by contrast, is less likely to receive broad distribution.

LinkedIn also added a “Seems like AI slop” option so users can flag posts or comments they perceive as empty AI-generated content.

During the first few weeks, more than one million people used that feature, according to figures released by LinkedIn and reported by Social Media Today.

The same publication says content LinkedIn internally classified as AI slop was getting around 40% fewer views than a few weeks earlier.

But correlation should not be confused with an automatic penalty.

LinkedIn has clarified that an individual “AI slop” report does not function as a policy violation and does not automatically cause a post to be downgraded for everyone. The signal primarily affects the reporting user’s experience and also helps LinkedIn understand what kind of content its community considers low value.

Here we find perhaps the clearest signal of the shift underway.

During the first years of the generative explosion, the incentive was:

produce more.

Now platforms are starting to ask:

produce more of what?

Substack turns AI detection into a reader tool

Substack has taken a different path.

Since July 2026 it has offered a feature called Scan for AI text, which uses Pangram technology to estimate what percentage of a text appears to have been written by a person and what percentage may have been generated or assisted by AI.

The tool works on posts and Notes published since July 21, 2026, and can also be used on comments and replies.

Substack also explains that neither the platform nor Pangram uses publishers’ content to train generative models.

That opens another inevitable debate.

AI detectors are not oracles.

They can become one additional signal for the reader, but mathematically determining where machine assistance ends and human authorship begins will become increasingly difficult as the two are mixed throughout the creative process.

A writer may personally research a story, build the argument, and then use a model to edit the prose.

Another may ask an AI to write the entire article and do little more than press “publish.”

Both used AI.

But they are radically different creative processes.

The real problem is not AI: it is removing the human from the process

The policies emerging across these platforms reveal a common trend.

The important divide will probably not be:

human content vs. AI-generated content.

It will be:

content with intent, judgment, and accountability vs. automated content with no added value.

A publication can use AI to research, organize information, correct grammar, or generate visual materials and still have a human thesis, investigation, and editorial decision behind it.

At the other extreme, it is possible to connect an API to a database, manufacture thousands of articles, videos, or posts, and distribute them automatically.

Both systems use the same technology.

But one uses AI as a multiplier of human work.

The other tries to make the human optional.

And that second model seems to be exactly the one starting to encounter resistance.

Content automation will have to evolve

This has especially interesting consequences for automated creation pipelines.

Until now, an obvious metric might have been how much content a system could produce:

10 articles per day
50 videos
500 posts

As models become cheaper and more capable, that metric loses importance.

The question becomes something else:

what makes each piece worth existing?

That forces automated systems to introduce things that, paradoxically, have little to do with generating more tokens:

  • editorial selection;
  • verifiable sources;
  • fact-checking;
  • original perspectives;
  • human or independent review;
  • criteria for discarding content;
  • transparency about AI use.

The best generative pipeline may not be the one that automatically publishes 100 pieces.

It may be the one that generates 100 candidates and has enough judgment to publish only five.

This connects with a broader transformation in agent engineering: as generation becomes cheaper, verification, selection, and governance become relatively more valuable.

Brands will have to explain how they use AI

Amy Zwagerman also points to another important consequence: organizations should start establishing their own public AI policies before platforms define public perception for them through a simple label.

A reasonable policy should explain:

  • where AI is used;
  • where it is not used;
  • which decisions remain under human supervision;
  • how private or confidential information is protected;
  • what customers should know about the creation process.

Because a label such as “Made with AI” conveys very little information.

It could mean that an image was artificially generated.

It could mean that a journalist used AI to organize twenty sources.

It could mean that an entire video was manufactured automatically.

Putting all of those processes into a single category will become less and less useful.

Useful transparency is not simply admitting that AI was involved.

It is explaining what the AI did and who retains responsibility for the result.

From “AI-generated” to “AI-assisted”

Perhaps we are approaching the end of the first stage of the generative Internet.

The stage of fascination with generating anything.

Text.

Images.

Videos.

Voices.

Avatars.

Everything automatically.

The next stage may be much more interesting: deciding what deserves to be automated and what should remain a person’s responsibility.

YouTube, LinkedIn, and Substack are sending different signals, but they all point in a similar direction.

Artificial intelligence is not disappearing from content creation.

Quite the opposite.

It will probably end up embedded in nearly every creative tool.

What is beginning to disappear is the idea that “AI-generated” is, by itself, a value proposition.

When anyone can generate infinite content, generation stops being the competitive advantage.

The advantage moves somewhere else again:

having something to say.

Sources