On AI Slop

Last month I shipped a feature to thredspan at half past eleven on a Tuesday night. A whole slice of a B2B SaaS product, specced, built, tested, deployed, standing up in production by midnight. I am the only person on it. There is no team. There is no standup the next morning where I report the work to anyone. There is just me, a working set of agents, and a product that did not exist a year ago.

There is something I am not allowed to say in polite engineering company. I am having more fun building software than I have had since my twenties. I left day-to-day development the better part of fifteen years ago because it had stopped being fun and started being a grind, and I did not expect to ever want it back. I have it back.

the slop is real, and I will not pretend otherwise

The case against AI is good, so I will make it first, in its strongest form.

The internet is filling up with rubbish faster than at any point I can remember. Auto-generated articles that say nothing, fake product reviews, AI-narrated videos over stolen footage, support bots that loop you back to the question you started with. Simon Willison popularised the word “slop” for this in May 2024, and he is careful to note he did not coin it. The word stuck because it named something everyone had started to feel. Stuff that was generated because it could be, and that nobody asked for.

The code version is worse, and the critics have a good receipt for it. Veracode’s Spring 2026 analysis found that only around 55% of AI-generated code passes basic security testing, a number that has stayed roughly flat for two years across GPT-5.x, Gemini 3 and Claude 4.x. Eighty-five percent of the samples failed to defend against cross-site scripting. That is not a rounding error. That is most of the code, most of the time, shipping a hole.

And the failures are not theoretical. A founder going by Leo watched his vibe-coded product get torn apart and posted “guys, i’m under attack… i’m not technical” while the attack was still happening. In 2026 a vibe-coded social network called Moltbook was breached within three days of launch, the reports pointing to a misconfigured database with no row-level security. I am hedging the exact details there because it is largely single-sourced, but the shape of it is real and I have seen the shape before.

So when an engineer tells me that a lot of what AI produces is slop, I do not argue. I agree. I am not here to tell you AI code is secure. I am telling you it often is not, and the receipts are right there.

That is the concession. Now for where I part company.

two arguments wearing one word

When people say “AI slop,” they are running two arguments at once, and the word lets them swap between the two without anyone noticing.

The first is a quality argument. It says: this output is bad, insecure, low-effort, and it is making the commons worse. That argument is true and I have just spent four paragraphs agreeing with it.

The second is a status argument. It says: people who should not be building software are now building software, and that is an affront. This one rarely gets stated out loud, because out loud it sounds like what it is. It hides inside the quality argument and borrows its credibility. The word “slop” does the swapping. It lets a status complaint travel in a quality complaint’s clothes.

You can feel the seam when you watch where the anger lands. Nobody is angry that a senior engineer with twenty years behind them uses a model to go faster. The anger is reserved for the person without the CS degree or the years behind them, who shipped a thing anyway. The complaint is dressed as “the code is bad.” The feeling underneath is “how dare you.”

I want to be fair to the people making the status argument, because they are not villains and some of them are excellent engineers. The fear is real and it is human. You spent years earning something hard, and now a tool hands a version of it to someone who did not pay the same price. That stings. I would feel it too. But a feeling being understandable is not the same as it being a reason, and we should not let the word “slop” smuggle the feeling in as though it were the reason.

I respect the price the engineers paid

The engineers were rare and expensive for a good reason. It was genuinely hard. I have real respect for the people who sat the AWS and Google Cloud exams, who learned what a VPC was and why it mattered, who could stand up a production deployment without it falling over the first time real traffic hit it. The knowledge needed to do basic things, a cloud deployment, a sensible database, a build pipeline that did not collapse, was deep and unforgiving. For fifteen years I watched it get harder, not easier, and I was glad to have moved out of the daily fight with it.

So when I say AI is democratising the production of software, I am not waving away the difficulty. A tool has arrived that handles a large part of it, which is the thing tools have always done. The respect for the engineers and the excitement about the democratisation are not in tension. The engineers were rare because the work was hard. The work got easier. Both sentences are true.

fix the claim before someone else does

The easy version of my own argument is a lie. Here is the lie, before anyone else gets to use it on me.

Building thredspan was not impossible for anyone a year ago. A funded team could have built it. Plenty did. The narrower claim is the true one: I could not have built it. Not alone. A year ago, building a complete B2B SaaS product by myself would have meant raising money, hiring a team, and burning months of runway before the first customer saw anything. That is the thing that changed. Software did not suddenly become possible. My software became possible, for me, without the apparatus that used to sit between an idea and a shipped product.

And the reason I keep coming back to that midnight deploy is that it is the most fun I have had in fifteen years. For most of the last fifteen years I was glad to be out of the code, because the code had become an argument with infrastructure that I usually lost. Now the argument is mostly handled, and what is left is the part I always loved: the value, the outcome, the thing in front of a user. The typing was never the joy. The typing was the toll you paid to reach it. The toll just got a lot cheaper.

name the gap the critics exploit

One strawman gets used to dismiss all of this. Name it once and it stops working.

Andrej Karpathy coined “vibe coding” and scoped it carefully when he did. His framing, in February 2025, was that it is “not too bad for throwaway weekend projects, but still quite amusing.” Throwaway. Weekend. That was the brief. Vibe coding, as the man who named it meant it, is for the disposable.

What I do on thredspan is not that. It is supervised building by someone who has shipped software for two decades, who reads every change, who knows what a missing authorisation check looks like, who tests the thing before it goes near a customer. The gatekeepers’ favourite move is to point at the weekend vibe-coder who shipped a breach, then act as though that is the whole democratisation argument. It is not. They are different activities. One of them produced the breach headlines. The other is quietly shipping working products. Hold the two apart and the status argument loses most of its ammunition.

the fear is sixty-seven years old

None of this is new. The fear that the wrong people will get to make the thing is one of the oldest fears in computing, and it predates every model by decades.

COBOL was designed in 1959 with an explicit goal that reads like it was written for this exact argument. The committee wanted a language readable enough that business managers could follow the programs themselves, and eventually write them, breaking the dependence on a small priesthood of specialists. The priesthood, naturally, was not thrilled. The idea that the people who own the problem might also write the solution, without a gatekeeper in the middle, was threatening in 1959 and it is threatening now. Sixty-seven years, same reflex.

Then look at what actually happened every time a tool widened the gate.

The spreadsheet is the cleanest case. VisiCalc shipped in 1979, and people who could now do their own numbers no longer needed a clerk to do them. You would expect that to gut the accounting profession. It did the opposite. Since 1980, by the Bureau of Labor Statistics figures that NPR’s Planet Money dug into, there are around 400,000 fewer bookkeeping and accounting clerks, and around 600,000 more accountants. The cheap part got automated. The judgement part grew.

The cash machine ran the same play against the same intuition. James Bessen’s much-cited work found that between 1985 and 2002, American ATMs went from around 60,000 to 352,000, and the number of bank tellers went up, from roughly 485,000 to 527,000. The machine made each branch cheaper to run, so banks opened more branches and hired more tellers, whose job shifted from counting cash to talking to customers. The thing that was supposed to end the job changed it and then needed more of it.

The panic itself is a tradition. Desktop publishing put typesetting on a desk in 1985, and the trade recoiled at the “ransom note effect,” all those amateurs with twelve fonts on one page. The newsletters were ugly for a while, then the taste caught up, and a whole profession of designers exists now that could not have existed when typesetting lived behind a guild. Kodak ran it even earlier. “You press the button, we do the rest,” 1888, and photography stopped being a chemist’s craft.

The pattern does not bend. A tool makes a hard thing easy, the people who owned the hard thing panic, the gate widens, the work multiplies and changes shape, and a few years later nobody can remember what the fuss was about.

the honest cost, including the part that hurts

The optimistic version skips the part that is true and unpleasant. The part it skips is the one that matters most.

Take the typewriter, which is the analogy people reach for and almost always get wrong. The comforting version says the typewriter created more office jobs than it destroyed, full stop, so relax. The sharper version is the one worth telling. The typewriter did grow the clerical workforce enormously between 1870 and 1930, and it did transform who did the work, with women going from a tiny fraction of clerical workers to more than half. But it killed a specific job, the copyist, the person who wrote documents out by hand, and it invented a new one, the typing pool, at lower pay and lower status. The workforce grew. And the people who lost the old job were not the people who got the new one. Both things are true at once, and any account that gives you only the happy half is selling you something.

That is the shape of what is happening now, and pretending otherwise would make me exactly the AI fanboy the critics accuse me of being.

The displacement is real and it is landing on the young. Stanford’s Canaries in the Coal Mine work, updated in November 2025, found roughly a 16% relative decline in employment for 22-to-25-year-olds in the occupations most exposed to AI. Entry-level software postings are down somewhere between a quarter and 40% from their 2022 peak, depending on whose count you read. The Pragmatic Engineer’s 2026 survey is among the more careful accounts of the same market. The juniors are the copyists in this story. They did nothing wrong, and the bottom rung of the ladder is being sawn off underneath them. I will not dress that up.

The productivity picture is murkier than either side admits. METR ran a randomised trial in July 2025 and found experienced developers were 19% slower with early-2025 tools while believing they were 20% faster. Their February 2026 follow-up hints the gap may be closing, but METR themselves call it an unreliable signal, so I will not claim a confirmed speedup. DORA’s 2025 report has the line that stays with me: AI amplifies whatever is already there. Good teams get better. Sloppy teams get sloppier, faster.

There is also a smarter version of the case against me than the status complaint, and I want to engage it rather than dodge it. An October 2025 paper on what its authors call the “slop economy” argues that democratisation has a class edge: the people who can pay get good, human-checked output, and the people who cannot get fobbed off with the AI-slop internet. That is the strongest objection to the cheerful frame, because it says the gate did not really open, it just moved, and the poor ended up on the wrong side of it again. I do not have a clean answer. I would only point at the spreadsheet and the cash machine and note that cheap tools have a long history of starting out as a worse product for the people at the bottom and ending up as the thing everyone uses. That is a bet, not a proof. I will come back to the bet.

slop is real, and it is also a moat

Satya Nadella has stood on both sides of this argument inside a year, and the contradiction is worth sitting with.

In January 2025 Nadella reached for Jevons paradox to explain AI’s future: “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” Jevons, writing about coal in 1865, observed that making a resource cheaper to use increases the total amount consumed rather than reducing it. Applied to software, the claim is that cheaper production means vastly more software gets made, not less, and the demand for everything around it grows to match.

Then through 2026 the same man spent his energy campaigning to get people to stop calling AI output “slop”. I find myself between Nadella and the purists, agreeing with neither all the way. The purists are right that slop is a real word for a real thing, and Nadella wanting it retired tells you exactly whose product gets called slop. But the good engineers manning the gate are doing Nadella’s opposite work for him. The word, in their mouths, is partly a moat. It keeps the production of software framed as something only the credentialed should attempt, and that framing has a status payoff for the people inside the gate.

Both things are true. The slop is real. The word is also doing gatekeeping work. You can hold both.

are we going to pretend the past was clean

There is one question the whole status argument cannot answer.

Are we going to pretend nobody wrote bad software before AI?

There was no golden age. Theodore Sturgeon, the science fiction writer, gave us the law for this in 1957, usually quoted as “ninety percent of everything is crud.” Sturgeon’s law, as it is popularly known, was not about software at all, but it has never been more applicable. Ninety percent of the code written by credentialed humans before any model existed was crud too. The spam was already there. The content farms were already there: Demand Media and eHow were industrialising low-value articles until Google’s Panda update in February 2011 knocked them down. We did not need AI to flood the commons with rubbish. We were perfectly capable of that ourselves.

AI did not invent slop. AI made slop cheaper to produce, the same way it made everything else cheaper to produce. That is a real problem about volume. It is not a new problem about kind. And “the volume of crud went up” is not, on its own, an argument that the wrong people are now allowed to build.

the bet, stated as a bet

The optimistic case is not a fact. It is a bet.

The bet, drawn from a couple of decades in and around consultancy, is that democratised production grows the total amount of work rather than shrinking it. More people able to build means more things built, means more things that need shaping, integrating, securing and maintaining, and explaining to the humans who have to live with them. In a consulting frame, every newly capable builder is a future client who has built something and now needs help making it good. I think this makes more consultants, more builders, more jobs, after a period of displacement that is real and painful and falling hardest on the people who can least afford it, the juniors, right now, this year.

I might be wrong. The slop-economy paper might be the truer read, and the juniors might not get a new rung to replace the one being sawn off. I am not certain. What I am is on the optimistic side of an uncertain bet, with my eyes open about the cost, because every previous version of this bet, the spreadsheet, the cash machine, the desktop, the camera, paid out the same way, and the people who bet against it spent the following decade explaining a fuss nobody else could remember.

And underneath all of it, the thing I keep not being able to say in polite engineering company. I am building again. At half past eleven on a Tuesday night, alone, with a product in front of me that a year ago would have needed a team and a round of funding I did not want to raise. The grind that drove me out of code fifteen years ago is mostly handled now, and what is left is the part I loved when I was twenty-two. I have not had this much fun since. If that makes me a fanboy, I have made my peace with the badge.