Why AI-Generated Crochet Patterns Don't Work: The Math
It usually starts with a beautiful photo. A dragon with perfect scales, a doll with an impossibly sweet face, a listing that looks like every other listing. The pattern arrives as a clean PDF, the first rounds go fine, and then somewhere around the head decreases the fabric stops making sense. The count does not close. The shape is not the shape. You frog it, try again, read the round a fifth time, and start wondering what you did wrong.
Often, nothing. The pattern was never workable, because of how it was made. And once you understand the mechanism, you can catch these patterns in seconds, with arithmetic instead of suspicion.
A crochet pattern is arithmetic wearing a disguise
Every round in an amigurumi pattern is a small equation. The stitches you work either keep the count, grow it, or shrink it. "(sc, inc) x 6" takes 12 stitches and produces 18. Six decreases take 12 stitches and produce 6. The count at the end of every round is not a stylistic choice, it is the consequence of the operations in that round, and the next round has to start from exactly what the last one left.
There is also a physical law hiding in the fabric. A round can at most double its stitch count, by working an increase into every single stitch, and can at most halve it, by working a decrease across every pair. A round of 22 stitches can become at most 44. It can never become 60, no matter what the text says, no matter how confident the formatting looks.
Human designers get this right because they crocheted the thing. The sample sat in their hands, the counts were counted, and when round 12 did not close they noticed, because they were holding round 12.
Why generated patterns break exactly here
Text generators write crochet patterns the way they write everything: by producing the next line that looks like what usually comes next. And a generated round genuinely looks right. It has a round number, plausible abbreviations, a tidy count in parentheses. What it does not have is a connection between the operations and the number, because the generator has no model of fabric, no running stitch count, and no idea that round 9 must consume what round 8 produced.
So the failures have a signature. Counts that are round and pretty, 30, 48, 60, instead of being the consequence of the operations written beside them. Jumps that no hook can perform. Rounds that consume stitches the previous round never made. The text reads fluently, because fluent text is exactly what the generator is good at. The arithmetic underneath does not hold, because arithmetic was never part of how the text was made.
This is worth sitting with for a second, because it means the problem is systematically detectable. You do not need to guess who or what wrote a pattern. You need to check whether the math works, and math is checkable.
Looking at the photos will not save you
The community's first instinct has been visual: guides on spotting suspicious images, extra fingers on hands, yarn that melts into itself, lighting that is a little too dreamy. Those guides are useful, and they are also fighting last year's war. Generated images improve every month, and a listing can pair a stolen photo of a real doll with generated instructions just as easily.
Authorship is the wrong question to stake your evening on. The right question is narrower and much easier to answer: can these instructions be worked as written? A machine-generated pattern with impossible counts fails that question. So does a human pattern with an unfixed typo in round 12, and those exist too, in large numbers. Checking the arithmetic protects you from both, without accusing anyone of anything.
How to check a pattern by hand
Three checks cover most of it, and they need nothing but a pencil.
First, recompute one round. Take any round with a repeat, multiply it out, and compare with the stated count. "(2 sc, inc) x 6" produces 24 stitches. If the line says 26, you have found something.
Second, follow the chain. The count each round states is what the next round has to work with. Six decreases need 12 stitches to consume. If the previous round left 8, the pattern is asking for stitches that do not exist.
Third, apply the doubling law. No round can more than double or less than halve the round before it. Any jump beyond that is not a difficult instruction, it is an impossible one.
Do this for two or three rounds in the middle of the pattern, where generated text tends to lose the thread, and you will know most of what you need to know before committing twenty hours of your life.
Or let the arithmetic run itself
We built the hand-check into a page, because nobody wants to recompute forty rounds with a pencil. Paste a pattern into the free pattern check and it recomputes every round, follows the chain between them, applies the doubling law, and flags the essentials a sellable pattern should carry, gauge, hook, yarn, assembly. Every finding shows its arithmetic and its line number, so you can verify the verdict yourself.
Two things about it matter here. It runs entirely in your browser, so the pattern you paste is not uploaded, stored, or seen by anyone, which matters when the pattern is someone's paid work. And it does not claim to know who wrote what. It reports that round 5 states 60 where 26 can grow to at most 52, and leaves the conclusions to you.
It cannot stop you from buying a broken pattern, nothing can, marketplaces do not allow tools between you and the checkout. What it changes is what a bad purchase costs you: five dollars and two minutes instead of five dollars and three evenings. And if you do end up requesting a refund, a report with line numbers and arithmetic is a much stronger message than "it seems wrong".
If you design patterns, run it on yourself
The same check is a designer's tool first. Run a draft through it before your testers see it, and they can spend their attention on wording, construction and feel, instead of catching a count a machine can catch. Inside the validation workflow that order matters: arithmetic first, humans second, because human attention is the expensive part.
In the Amigurumake studio the check runs automatically on every imported pattern, so nothing with a broken count slips quietly into a catalog. Honest designers have the least to fear from arithmetic and the most to gain from being checkable: when buyers learn that counts can be verified, verified counts become a reason to buy.
The frustrating evening at the top of this article was never about one bad pattern. It was about not being able to tell in advance. That part, at least, is now solvable in seconds.