We fact-check our own marketing. Here is what it caught this week

What slop actually is

Slop is not "AI-written content." Slop is unverified content presented confidently. The reason AI makes it dangerous is counterintuitive: AI research output is more polished and more sourced-looking than the average human first draft, which makes it more likely to be passed along, not less. A confident paragraph with five citations gets forwarded. Nobody opens the five links.

That is the trap. The citations are the camouflage. A claim feels checked because it has a footnote, and the footnote is exactly the thing nobody checks.

There is a second reason, and it is about us, not the machine. We adopted AI faster than we learned to use it. People treat it like Google, a lookup that returns facts that exist somewhere, when it is a generator that produces fluent, confident text whether or not the underlying fact is real. The training did not arrive with the tool. A search-engine mental model pointed at a fabrication engine is exactly where the damage comes from.

The mechanism

Our rule is simple and old-fashioned: every metric, every claim, and every citation is treated as false until it traces to a real, credible source we can open and read. In practice that means pulling each claim out of the draft and verifying it on its own, against the primary source where possible. We rank sources by credibility and throw out the weak ones. We read the studies, not the summaries of the studies. And the part that matters most: anything that cannot be traced gets cut, not softened. We do not rewrite an unverifiable claim into a vaguer version that is harder to falsify. We delete it.

It is not clever and it does not require special tooling. It requires being willing to lose material you liked. Most of the discipline is emotional, not technical: the draft reads well, the stat is perfect for your argument, and you have to cut it anyway because it does not check out.

What it caught this week

Here is what the pass actually removed or corrected from this week's raw material. None of these were obvious. All of them looked fine.

A headline thesis that no one had actually said. The summary opened on a confident framing attributed to a major bank, that profit-margin expectations had "decoupled from fundamentals." When we went looking for the note that said it, there was no such note. The AI had synthesized a plausible-sounding thesis out of separate, real fragments and attached an authoritative name to it. Cut.

A real statistic attached to the wrong study. A "20 percent" figure was presented as proof that one thing caused another. The 20 percent was real, but the study it came from was measuring something different entirely. The number was true and the claim built on it was false. Corrected, and the claim dropped.

An academic paper cited for the opposite of what it found. This is the one that should worry you. The draft cited a specific arXiv paper as evidence that AI productivity gains are smaller than people believe. We pulled the paper and read it. It reports the opposite: meaningful positive gains. Had we trusted the citation, we would have published a paper as supporting our point while it argued against it. Dropped entirely.

A vivid analogy that belonged to someone else and meant something else. A striking comparison was attributed to "the market." It actually came from one named economist, and it was about a different subject than the draft used it for. Cut.

A stack of citations no one should publish next to. Several "sources" were crypto-news aggregators, a content farm, and an AI-generated summary page. Every one was replaced with the primary source or removed.

We are not naming the tool that produced the summary, because the point is not that one tool is bad. This is what AI research output does, routinely, across tools. The polish is real. The reliability is not. The two are not related.

This is already costing the people who skip it

Our examples were low-stakes: a marketing draft, caught before anyone saw it. The same failure, unchecked, is already producing real consequences, and the courts are the clearest record of it.

In 2023, a New York lawyer filed a brief written with ChatGPT in Mata v. Avianca. It cited cases that did not exist, complete with quotes attributed to real judges who never wrote them. The judge sanctioned the lawyers and the story went around the world. The instructive part is what happened next: it kept happening. In 2025, attorneys for MyPillow's Mike Lindell were fined after a brief came in with roughly 30 defective citations, including cases that do not exist, and the same lawyer was sanctioned again in 2026. These are not anonymous. They are named lawyers, in public filings, in front of judges who now go looking for it.

And it is not isolated. A legal researcher has been cataloguing court decisions that call out AI-fabricated content, and the count already runs into the hundreds, the large majority of them in the last year. Judges have gone from surprised to expectant.

It is not just lawyers. A Canadian tribunal held Air Canada liable when its website chatbot invented a refund policy that did not exist, and the airline's argument that the chatbot was "a separate legal entity responsible for its own actions" did not survive contact with the adjudicator. The fabrication reached a customer, and the company owned it.

And it reaches our own corner of the world. In 2025, Deloitte's Australian arm delivered a government report that contained references to academic papers that did not exist and a fabricated quote from a court judgment, and refunded part of the fee once a researcher caught it. The uncomfortable lesson is that this is not a small-shop or junior-analyst problem. It happens anywhere the verification step is assumed instead of run, including at the largest firms.

The common thread in all of it: smart, capable people, under deadline, treated a confident answer as a checked one. Not one of these required a bad actor. It required a skipped step.

Why this is the difference between sharing and spreading

Sharing information has a verification step. Spreading slop skips it. That is the entire distinction, and it is mechanical, not moral. You do not avoid spreading slop by being a good person who would never lie. You avoid it by running a step that catches the confident falsehoods you would otherwise have repeated in good faith. Good intentions are exactly how slop spreads, because the person forwarding it believes it.

The cost of skipping the step is not abstract. In a technical or regulated audience, one fabricated statistic is the end of your credibility for the whole piece. The reader who catches the bad citation stops trusting the other twelve. We would rather cut a great-sounding stat than be the firm that got caught citing a paper that said the opposite.

The standard, inside and out

This is the bar for everything we publish, and we hold it internally before anyone outside sees a word. It is also why the series this came from is worth trusting: not because we say so, but because we can show you what we removed to get there. And it is a mechanism you can adopt today, no tooling required: pull every claim, open every source, rank them, read the primary, and cut what will not trace. The discipline is the product.

One-sentence takeaway

The line between sharing information and spreading slop is not who or what wrote it, it is whether a verification step ran before it shipped, and we run that step on our own work and can show you what it cut.

What this is not

This is not an argument against using AI. We use it heavily, including to run the verification pass itself. It is not a claim that our process is perfect, a mechanism reduces error, it does not eliminate it, and we will miss things. And it is not a claim that all AI output is slop. The good stuff and the slop look identical until someone checks. The checking is the only difference.