In Defense of Slop

"Slop" is what you call AI writing when you want to stop reading it. Voice-less, sourceless, confidently wrong, churned out to fill a hole where a person used to sit. The contempt is earned. I share it.

But it's aimed at the wrong thing. The problem was never that a machine helped. The problem is who the writing targets, how carelessly it's built, whether anyone signs it, whether you can check it, and whether the person who can publish a thing knows when not to. Get those five right and the same tools produce the opposite of slop — the local accountability reporting that few newsrooms still staff for, and that otherwise just doesn't happen.

So I want to defend slop. Not all of it. Slop under five conditions.

One: punch up, not down

The slop that earns the sneer punches down. It puts generated copy under a fake byline where a paid writer used to be. It floods a content farm for the ad pennies a person used to earn.

The slop I'll defend points the same tools at institutions that have run for decades without anyone looking — and produces work that, without the tools, wouldn't exist. Not because the work stopped mattering. Because the jobs that did it have largely gone: the research librarian who knew where the documents were, the fact-checker who ran down every figure, the three paid months to follow one thread. The records are still public. What's new is that one person, careful, can read them at the speed a whole desk used to.

In March I published Olive Branch, about one empty lot — 920 Olive Way, which the district bought for $56.4 million and has let sit fenced and dead ever since. To explain why a fully entitled tower site produces nothing, I had to assemble the story out of six public-record systems that have nothing to do with each other — assessor, recorder, audited financials, board minutes, a bond statement, SEC filings — plus a handful of open-web sources to tie them together, down to the archived page holding the renderings the district had quietly pulled offline. No one of them is the story. The assembly is the story.

Here's the before and after. That assembly was the work of the desk that no longer exists. What the machine does is put it back within reach of one person: it finds the documents and pulls the passages, I decide what they mean and I sign it. The piece exists only because the tool made that survivable alone. Not that the work got easier — that without the tool it simply doesn't happen.

It also surfaces what no one else is watching. In May I found a facility going up inside the convention center — a "public safety hub" called C-STAR, completed two weeks before the World Cup lands in Seattle — with no public account anywhere. I wrote it up, then put the question to the board in person: does anything keep immigration enforcement off PFD property? The answer came back in writing — "no such policy has been adopted by the PFD Board". A records request shook loose the rendered drawings of what it looks like, which the district doesn't publish. No newsroom was on it. One resident with public records was.

Slop that fills the vacuum left by collapsed institutions is doing public work. Slop that creates the vacuum is doing private harm.

Two: built one claim at a time

The slop that earns the name is the one-shot dump — prompt in, paragraphs out, ship it, a human "reviewer" rubber-stamping the top. It never survives contact with a source because it never touched one.

I work the other way. Every claim that carries weight is its own unit: sourced, drafted, checked, before it's allowed to stand next to the others. The machine is trusted to find and to draft, never to be the last word on what's true. I spot-check the figures it recites rather than re-deriving every one, bank what's confirmed where I can reuse it, and build only on what I can cite. I read and re-read every piece before it's published, looking for places to improve. Drafting is close to instantaneous; editing remains an iterative exercise in close reading, and it's still the bottleneck. Where that line falls — between what I verify and what I take on trust — is something I keep honest about, because claiming I'd checked every word would be its own kind of slop.

Take the bond debt behind Summit 2030: Big Balloon Bailout — I doubt I'd have found the cliff without the machine. The official statements sit on EMMA, long filings thick with debt-service schedules, and the number that matters is buried in the tables: the annual payments nearly double at the end of the decade. Lining those years up by hand is an afternoon of cross-referencing; the machine did it in a query. Finding it was the machine's job; deciding it was true was mine — I opened the official statements and read the schedules, because a claim that consequential doesn't get to stand on the machine's say-so.

What it looks like on the page: when I wrote up C-STAR, nearly every sentence ended in a citation — the square footage, the sealed officers' room, the headcounts, each bolted to a specific schedule or section of the bid documents, quoted rather than paraphrased. Built one claim at a time still reads like a smooth essay — it should, for a human reader and an AI agent alike. What's different is underneath: each claim that carries weight is bolted to a document you can open yourself, so none of the smoothness asks to be taken on trust.

Three: signed, with corrections

Anonymous slop answers to no one. Wrong into the void; nobody owns it, so nobody fixes it. A byline is a promise: I made this, I stand behind it, and when I get it wrong I'll say so in public, in the same place I said the wrong thing — dated, visible, not a quiet edit pretending it never happened.

Putting my name on it adds necessary friction that makes this slower than it could be. And even so, I move fast. Old-school journalism confirmed everything before it ran, because the printed page was the record and the correction came later, buried, where it never caught up. On the web the correction lives in the same place as the claim, dated, the moment I know — which changes what I can responsibly publish. A fact still has to be sourced before it stands. But a conclusion is different: if two public filings point somewhere, I don't have to walk each one past the agency before I say where they point — I can publish the inference, marked as an inference, and correct it in the open if I'm wrong. Fast to generate, fast to publish, fast to correct. The speed is warranted only because the correction is as visible as the claim.

That's not theoretical. In an op-ed in The Urbanist, I claimed the convention center's older building was "dark" most of the year — an inference drawn off the public event calendar, before I had the booking data. The CEO wrote to say my numbers were off. So I filed a records request for the actual booking data, and the agency's own records contradicted me. I printed the correction in the original article and did an in-depth dive into the numbers in my next piece. You don't get to point this at a powerful institution unless you're willing to be corrected by it in public.

Smaller ones, too. I misread a floor-plan note in the C-STAR piece, and an architect who reads the Dispatch wrote in with the fix; I posted it.

And signing it is more than owning the errors. It's carrying the thing to the people who can act on it, under your name, and eating whatever that costs — because publishing into a void and walking off is the tell of someone who doesn't quite believe their own findings. This was not ship-and-forget. I went to a board meeting and asked my question out loud, in the room. I sent named inquiries — not anonymous tips — to the district's executives, the Mayor's office, the King County Council, city councilmembers, reporters whose work I trust. I took it to people I know, people I was sure would care. Most of it went nowhere. That exposure is the cost of the byline. It's also the point of it. A name on the work is a standing bet that it's worth the awkwardness of being the guy who won't let it go.

Four: transparent, on a spectrum

You can check the work.

Show your sources — mandatory. Every piece carries a source list; every claim that matters traces to a document you can open. This is the floor. Below it there's no defense, just assertion wearing footnotes. "AI told me" is not a source. If you take one thing from this, take that.

Show your code — optional, ideal. Not the throwaway script behind any one piece — that you can regenerate with AI in minutes, so there's little point posting it. The code worth publishing is the reusable kind: tools built so the next person can run the same sort of inquiry on their own targets. pdc-mcp is mine — an open-source server that wraps the campaign-contribution and expenditure records Washington's Public Disclosure Commission publishes, so anyone can point their own AI at them and ask who funded whom. Not the script I ran for one story; a tool you can run for yours.

Show your prompts — not needed, but why not. The prompts aren't the proof; the sources are. Nobody needs my wording to check a number against the audited statements. But there's no reason to hide them, and showing them lowers the bar for the next person.

The more of the process I expose, the less my word matters — which is the right direction. And transparency only counts if it includes the soft spots. When I searched the federal databases for a contract behind C-STAR and found none, I handled the absence carefully, because the district's own federal-contractor registration had lapsed in 2024 — so "no record" is consistent with no federal involvement and with involvement routed some other way. That search also made me walk back an earlier claim: I'd first called C-STAR a "federal coordination facility," and when the records wouldn't carry the word federal, I corrected it to "public safety hub" and showed why. Slop you can fully audit — sources, code, and the weak spots flagged — stops being slop.

The deepest soft spot sits upstream of all that checking. When the machine chooses which documents to hand back, it makes an editorial call before I see a single page, and that call can fail two ways: it can surface something that shouldn't carry weight, or bury something that should. The first kind I can usually catch — whatever it hands me is right there to read in full and weigh against the rest. The second is the one I trust least: I can confirm every figure in every document it surfaces and still never learn about the one it left out. Say I ask for the records on a single parcel: back come the audit, the bond statement, the assessor file, and I check each against its source. What the machine can't flag is the record it never matched — a side agreement filed under the developer's name, a resolution that names the lot only by number — which might be the very document that turns the story. I guard against it where I can — reading the whole document, not just the excerpt the machine pulled, and hunting for the fact that would break my story rather than the ones that confirm it. I don't manage that on every claim; the more weight a claim carries, the harder I look. But what it leaves out is the hardest error to catch, and the guard is only ever partial.

The same toolchain that reads records also draws pictures, and the line between the two has to stay visible. The bond math behind Summit 2030 — the cliff — traces straight to the EMMA filings, deterministic and reproducible. The image at the top of that piece, the Summit swollen into a balloon, came off an image model and proves nothing on its own. That's why its caption says "AI visualization" in plain sight: a reader is owed the difference between a number I can trace to a filing and a cartoon I made to carry the idea. The discipline that traces a figure to its footnote has to extend to the pictures — marking, every time, which strokes on the page are evidence and which are only there to help you see.

Five: disclose for a reason, not because you can

A public record I can lawfully pull is something I may publish. That alone never made it something I should. The records act answers "may I." It says nothing about "should I," and "I had it" was never a reason to put anything in front of anyone. Seek the truth and minimize the harm — every newsroom code says some version.

When I asked, in print, what C-STAR was — a public-safety facility with immigration-enforcement implications, opening in a public building, its users unnamed — and the district wouldn't answer, publishing what I had was the only way to get a question on the table the public had every right to ask. The disclosure was the remedy for the dodge. The campus master-plan drawings are a different animal. Nobody dodged a question; I came across draft design work for a plan that isn't finished. The accountability questions in it are real — whether it's a smart use of public money to spend heavily sliding an escalator over to grow the retail floor, whether a neighborhood's plan ought to be drawn with only its landowners in the room — but I can make every one of those by writing about them and showing the one slide each turns on. I don't need to post thirty-five renderings to say the escalator math is upside down. Disclose what the argument needs. Not one slide more. The few I do show — credited, used to report a question the public is owed — I'm claiming as fair use: limited, attributed, news and criticism.

I have to show every source I use. The institution discloses what the law strictly requires, and within that it does the job well: thin minutes, records requests filled promptly and accurately. What the old order never did was press it past the minimum — a certain kind of paternalism, supported by a quiescent oversight mechanism. Work like this supplies the pressure that was missing, reasons to open up that didn't exist before: a wrap on the C-STAR window describing what it was going to be as soon as it got the green light; financial statements that put the real loss up front instead of buried in a footnote.

So weigh it, risk against reward, because the records channel is a commons and it moves both ways. Over-disclose — dump everything you can get for the sake of getting it — and the lesson the institution takes is to quit putting anything real in board materials, or to go lobby the rules narrower, and the channel shuts for everyone behind you. But the commons widens, too, and often it should: disclosure that surfaces a real public-interest question makes secrecy more expensive and openness more ordinary, and that's worth pushing hard for. So the question is never "can I?" It's "what does the public gain, and does publishing this pry the opening wider or get it nailed shut?" When the gain is real, you push. When it isn't, restraint is what keeps the window open for the next request.

None of which is deference. The discipline is about how much I publish and when, not what I'll touch. I live here; my money is heavily in Seattle real estate, not cash, so I'm tied to this city's success by more than sentiment — which is exactly why I won't knife its competitive position for a scoop, and exactly why I'll still write that the plan for my own neighborhood got drawn without the neighborhood in the room. "I could publish this" is where the call starts.

The mistake the method is built on

The lesson came the expensive way. Early on, I claimed the convention center's food-and-beverage contract with Aramark expired in January 2027. A whole framework sat on it — I'd built the case around a "2027 hinge," the year the district's options would reopen — and the date was already in my models and in emails I'd sent to people. It was wrong. No such expiration: Aramark had signed a fresh contract in 2025 that runs as far as 2033. Worse, when I went looking for where "January 2027" came from, I couldn't find a source at all. It had shown up in my notes with no provenance — most likely the machine's own, a confident detail it produced that I never traced, exactly what happens when you let the draft get ahead of the documents. And I found out the worst possible way: at a public board meeting, an Aramark executive stated the real contract term out loud — and it wasn't the date I'd published. No one set out to correct me; a number was simply said in a room, and it didn't match my own site. One number — unsourced, falsified in the open — and it took the whole framework down with it.

That failure is the before, not the method — the origin of the discipline, not an example of it. But "before" names a class of mistake, not a trapdoor for the next one: a claim that's sourced, sitting in the registry, and still wrong would be the method's to answer for, and I'd owe it the same public correction this one got. There was no registry then and no source-tracing habit; there was a confident number with no provenance and a framework built on top of it. Everything in this essay grew in the ground that mistake burned clear: every claim now has to trace to a primary source before it stands, and I keep a registry tying each one to a document, a date, a page. The whole correction is posted, permanently, at commons.conventioncityseattle.com/correction — because I'd rather you read my worst mistake than take my word that I don't make them. And what grew there isn't finished. This is the current method, not the final one — sharper than a year ago, surely clumsier than a year from now. A discipline that stopped changing would be the truest sign it had curdled into slop.

Did the discipline take? Look at the distance between that failure and the most recent one. While I was drafting this essay, I re-ran the published utilization numbers against the original spreadsheet, just to be sure they still held. Ten of eleven reproduced exactly. One didn't — a single figure for the older building's 2025 activity that I'd carried over from a working draft and never reconciled against the final data. So I corrected it, in public, with a dated note. The difference between the two is the whole story: the Aramark date was unsourced, and it surfaced in public — the real term said aloud at a board meeting, against what I'd published; this one was sourced and reproducible, and I caught it myself, before anyone else could. Re-running your own numbers from the source is how you find the one that slipped.

The word "slop" was never really about the machine. It was about the care — whether anybody took any. Aimed up instead of down, built one sourced claim at a time, signed by someone who'll answer for it, shown plainly enough that you can check every step, and disclosed for a reason instead of for the thrill of having it — that's not what the word was coined to insult. That's just journalism, done by one person, with a new tool, at a scale that used to take a newsroom.

What I'd ask in return is the right standard for reading it. Not the benefit of the doubt — credit is earned, not extended on request — but a fair test: whether the claims are sourced, whether the analysis reproduces, whether the byline answers for its mistakes. The value of a piece like this is in the curation, not the word choices — in which records got pulled and how carefully they were read, not in whether the sentences have the right literary pedigree. A polished hit piece and a sourced one can read exactly alike on the surface; what separates them is whether you can check the work, and the whole point of the method is that you can.

And measure it against the right thing — which is not perfection. No journalism is perfect; this one won't be the exception, and I've spent the essay marking where it can fail. The test is comparative: hold this against the two things that actually exist — the coverage this subject gets elsewhere, and the coverage it gets nowhere at all. The dead lot, the bond cliff, C-STAR: no newsroom was on them, so the bar is only whether one person with these tools put more on the record, better sourced, than would exist otherwise. The aim was never to be flawless. It was to widen what a single reader of public records can hold to account.

The data is public. It always was. The only question is whether anyone's reading it — and whether they'll put their name on what they find.

A note on how this was made, since the essay demands it of everyone else. I drafted it with the same AI tooling it describes, working from a month of my own reporting, then checked it against its sources and rewrote it in my own voice. That last step was most of the work: the machine writes in tells — the tidy summation, the little signpost that announces the next move, the balanced aphorism that sounds like a point and lands on none — and a draft is honest only once those are hunted down and cut. Stripping the machine out of the prose is the edit a reader can't do for themselves, so it's the one I owe them. Then I signed it. And the sources are here: every case links to the piece it's drawn from, each carrying its own sourcing — the argument rests on that record, not on my say-so. This is the disclosure I argue every outlet owes its readers — consider it the Dispatch's, with the standing version on our How the Dispatch Is Made page. It also burns real compute; I think the public benefit clears that cost, and I know not everyone will agree.

And I tested it the way I'm asking you to. I ran this essay through the read-it-with-your-own-AI tool at the top of the page, asked it for the strongest case against me, and rewrote against what came back — the comparative standard a few paragraphs up is one of the repairs, and so are several of the limits I now own instead of gloss. Turning the tool on the essay that describes the tool is dizzying, I know, and I'm sorry if it's hard to hold on a first read. But running this loop myself has a ceiling — the machine can only argue against its own work so far. The real test is a reader who knows something I don't: someone who can push the critique past where the AI could take it, turn up the support it missed, or read the whole thing and still change their mind. That's what the tool at the top is for — not just me sharpening my own work, but your frame entering where mine couldn't reach.

Ivan Schneider is the founding editor of The Convention City Dispatch.