Introducing the Gyrusfisher

The views expressed here are solely my own. They do not represent the opinions, positions, or policies of any current or former employer, client, or affiliated organization.

I built an AI tool to measure the level of structural sloppiness in a piece of content. Most platforms try to solve the "slop" crisis by measuring who wrote the words or how readers react to the vibe. They are looking at the wrong signals. Gyrusfisher ignores statistical sentence patterns and crowdsourced complaints. Instead, it analyzes whether the underlying architecture of your argument already exists elsewhere on the web. It is a diagnostic mirror for composition, not a post-publication filter.

Two days after I shipped the first version of this tool, LinkedIn rolled out a "Seems like AI slop" button. Users can now flag a post from the three-dot menu, hiding it from their feed and feeding the signal back into LinkedIn's classifiers.

Their chief product officer admitted that slop is hard to define and that the definition keeps changing.

That admission is honest, but it confirms a fundamental diagnostic error. LinkedIn's button asks people to vote on a vibe. It crowdsources suspicion instead of measuring structure.

A user clicks the button because a post feels generic, hollow, or recycled. That instinct is usually right. But the system built around that instinct still cannot tell you why.

It cannot distinguish a post that recycles a tired list of "5 leadership lessons" from a post that happens to share a familiar topic but says something nobody has said quite that way before. It measures reaction, not composition.

Other platforms make a different version of the same mistake. Substack attempted to solve this with a classifier that detects statistical fingerprints of machine output.

Both approaches suffer from the same tool-first trap. They measure proxies for the thing that actually matters, and neither one is measuring whether the content contributes anything new.

Content shock is not a product of the generative AI era. It predates the transformer by decades.

Scholars complained about the overwhelming volume of worthless books within a single century of the printing press. In the 1970s, Nobel laureate Herbert Simon was already describing information overload, watching corporate executives drown in paper memos.

The lazy publisher has always existed. What generative AI did was remove the friction.

It made producing the same tired argument in the same tired order nearly free. The volume of that particular failure mode exploded because the economic guardrails fell away.

AI did not invent slop. It gave slop a printing press with no ink costs.

Which is exactly why detecting "AI involvement" was always going to be the wrong axis. A human being can write pure slop using a rigid template and zero original thought.

Conversely, a writer working closely with a large language model can produce something genuinely new. The variable that matters was never who typed the words.

It is whether the specific combination of ideas already existed before this version showed up. It is about the presence of a gyrusprint—the structural signal of a person's judgment, criteria, and point of view.

I built Gyrusfisher to test this thesis. It is a prototype running on top of Claude's infrastructure, built in a single afternoon, but its logic is precise.

You paste a piece of writing in, and the tool does not ask whether AI touched it. It asks a much sharper question: does the full structure of your argument, the synthesis, the sequence, and the conclusion, already exist somewhere else on the internet?

If you paste in a draft that copies a standard industry template, the tool flags it. Not because an LLM wrote it, but because the intellectual work was never done.

This distinction is the entire point. A tool that cannot tell the difference between citing Herbert Simon and copying someone's entire argument about Herbert Simon is not measuring originality. It is measuring nervousness.

Every major platform treats slop as a moderation problem. They want to detect the bad thing after it is published, then hide it, flag it, or badge it.

I think that is backwards. The actual fix sits earlier, at the point of composition, not the point of consumption.

If a writer could see, before hitting publish, that their argument's structure already exists somewhere else with the same sequence and the same punchline, most of them would go back and find their own angle.

They would do this not because a platform threatened to punish them, but because no serious writer wants to publish something that has already been said.

Gyrusfisher is not a moderation tool. It is closer to a mirror you hold up to your own draft before anyone else sees it.

I am not pretending this single afternoon's build solves serialized slop for the entire internet. It is a prototype, and I am still working out where it lives permanently.

But it is a working demonstration that the direction is right. We must measure the structure of the argument, not the author, not the vibe, and not the statistical fingerprint of the sentence.

We do not need better filters to hide bad writing after the fact. We need better mirrors to expose empty thinking before it is published.

The quality of content is not determined by the hand that typed it, but by the mind that structured it. If you want to escape the slop cycle, stop measuring the tool. Measure the architecture.

Juan Carlos Vásquez has spent ten years inside enterprise content operations, repairing content supply chains before scaling them. Fix to Flow is the discipline that work produced. The views here are his own.