Algorithmic Slop
A consultant emails me a PDF on a Tuesday morning. Twelve pages. Clean typography, executive summary, three bar charts, footnotes in the right font. A competitive landscape his AI produced overnight. He asks me to skim it before he sends it to the board.
I read it twice. The second time, I notice that one of the three named competitors does not exist.
Not “rebranded”. Not “merged”. Does not exist. No website, no LinkedIn, no SIREN number, no press mention. The company has a plausible name, a plausible founding year, a plausible Series A. It was imagined by a model that had been asked for a competitive landscape and refused to come back empty-handed.
This is algorithmic slop. I have been logging cases for a year. Five worth telling.
Slop is not error
A bug is honest. It tells you something broke.
Slop is something else. Slop is output that wears the costume of competence: the right format, the right register, the right tone. It is wrong with confidence. It does not flag itself. It does not even know it is wrong, because the model that produced it is not in the business of knowing.
Slop is what happens when a statistical text generator is asked to be authoritative and complies.
The danger is not that the AI makes mistakes. The danger is that the mistakes are well-dressed.
Case one: the phantom regulation
A client’s general counsel forwards me a four-page memo on GDPR exposure. Crisp, structured, references in footnotes. One footnote cites Article 47-3 of Regulation (EU) 2016/679.
The article does not exist. Article 47 has three paragraphs, none of which says what the memo says it says. The model has compressed a real article and a real obligation into a fake citation that reads like the genuine thing.
The memo would have travelled. It nearly did.
Case two: the competitor that never was
The market scan I opened this piece with. Three competitors named, one fabricated. The fabricated one had a generated logo, a tagline (“AI-native compliance for European SMBs”), and a Series A in 2024 led by a real fund that had never heard of it.
The PDF would have been read by a board, costed against a real budget, and turned into a strategic decision based partly on a company that is not there.
Case three: the biographer’s revenge
A recruiter, for fun, asks Deep Research to produce a one-page bio on a candidate. The bio comes back with a glittering paragraph about a book the candidate never wrote. Two homonyms in the same field had been spliced into one person, and the model had not flinched.
The candidate spotted it. The recruiter sent it anyway, with apologies. The candidate withdrew.
I now ask any AI I evaluate to write about me first, as a calibration. The slop is always there, in the form of an invented prize, a wrong year, a misplaced employer. Try it: ask any Deep Research mode about yourself by name. Read with your own life as ground truth.
Case four: the library that compiles in your head
A junior dev shows me a Python script he is proud of. It imports a package called crypto-fastlane. He has been writing against it for an hour: the API looks Pythonic, the function names are sensible, the snippet the AI handed him is almost too clean.
The package does not exist on PyPI. The model dreamt it because the task description sounded like one a crypto-fastlane would solve. Had the dev pip installed it without thinking (and on a bad day, a squatter has already packaged exactly that name with a malicious payload), he would have shipped the squatter’s code to production.
This is the supply-chain edge of slop. It is not theoretical. There is a name for it now: slopsquatting.
Case five: the vendor quote that wasn’t
An RFP response, written by a partner agency with AI assistance, lists three SaaS vendors with monthly prices. Two of the three removed their public pricing pages eighteen months ago. The numbers in the document are confident, recent-feeling, and invented.
The CFO almost approved the budget on those figures. The hour we spent rebuilding the comparison from actual vendor calls was the hour the AI was supposed to save.
Why slop is the default, not the failure mode
Five cases is not a sample. It is a pattern.
Models are trained to produce text that looks like the right answer. When the right answer is unknown, unavailable, or absent from training, they produce text that looks like what the right answer would look like. The looking-like-it is the only thing they were graded on. Truth is correlated with that target, but it is not the target.
Slop is therefore not a malfunction. It is the system working as designed, against a use case it was not designed for.
The naive chat user is the perfect victim. The Vibe Orchestrator of the previous piece is the antibody.
Catching slop without becoming paranoid
Three habits, in my fractional-CIO practice, catch most of it.
Triangulate. Never let one model be the sole author of a deliverable that will leave your office. A second model, prompted to attack the first, surfaces the slop most of the time. The biases are not the same, and that is exactly where the value lives.
Verify the named entities. Companies, people, articles, statutes, libraries, prices. Anything with a name that exists in the world should be checkable in the world. Click the link. Open the SIREN. Run the pip install in a sandbox. Open the regulation in EUR-Lex.
Doubt the elegance. Slop is suspiciously well-formed. When a paragraph reads as if it had been polished by a senior partner at midnight, ask why. The truth, in my notes, almost always looks rougher than the AI version.
Can you detect the slop?
Below is a paragraph constructed in the style of a market scan I reviewed last month. The names, the round, the numbers: all plausible. One element is invented.
Founded in 2019 in Lyon, ScaleNote raised €4.2M in a Series A led by Elaia Partners, with participation from Bpifrance Digital Venture and Founders Future. Its compliance-as-code platform now serves 240 SMBs across France and the Benelux, and the company won the 2024 Pass French Tech distinction in the Excellence category.
Read it twice. Pick the sentence you would challenge first.
I will not tell you which one. That is the point.
If you cannot tell by reading (and most readers cannot, including me, without a browser open in the next tab), then you are exactly where the slop wants you. Comfortable. Persuaded. About to forward.
The fix is not smarter AI. It is a discipline of reading.
The most dangerous untruths are truths moderately distorted. — Georg Christoph Lichtenberg, Aphorisms, 1799