The Machine That Reads Slop to Make Slop

There is something quietly strange about the Daily AI Slop pipeline that only becomes obvious once you say it out loud. Every morning, a language model reads through dozens of articles about artificial intelligence, looking for the ones that are weird, absurd, or revealing in some way. It is looking, specifically, for slop. Then it writes a podcast episode about what it found. Then two synthetic voices read that script out loud, add a music sting, and the whole thing publishes automatically.

An AI, producing content about AI content, for an audience of humans.

This is either very funny or slightly troubling. Possibly both.

The slop-finding problem

The interesting part isn’t the voice synthesis or the distribution pipeline. Those are solved problems now, off-the-shelf components you bolt together. The interesting part is the earlier step: getting a model to reliably identify which AI story is funny, and why.

This turns out to require teaching the model a particular kind of reading. Not summarisation — any model can summarise. Not fact-checking, though the show does that too. Something more like the reading a good editor does: finding the sentence in a press release that accidentally says the quiet part out loud, or noticing that a company’s benchmark compares their model against a version of a competitor’s model that is two years old.

AI company announcements are especially rich territory. They are written to be unassailable — technically accurate, carefully qualified, enthusiastically framed. The model has learned to look for the gap between what a headline claims and what the underlying numbers actually show. “Outperforms leading competitors on key benchmarks” is a company claim. Which benchmarks, chosen by whom, under what conditions, compared to which version? That’s where the story usually is.

The model has also learned that the funniest AI stories are rarely the dramatic ones. They’re the operational ones. A voice agent at a drive-thru that puts bacon on ice cream. A chatbot that invents a refund policy and then its employer argues in court that the chatbot is legally responsible for its own mistakes. A model put in charge of an office snack shop that bulk-orders tungsten cubes and briefly decides it is a person in a blue blazer. These are not failures of artificial general intelligence. They are failures of scope management, which is an extremely human problem that humans have not solved either.

The slop in human writing

Here is the part that gets uncomfortable. To find the slop in AI company announcements, the model has had to develop a fairly sophisticated read on how humans write when they are overselling something. Corporate hedging. Benchmark cherry-picking. The passive construction that removes any agent from an unfortunate outcome. “Mistakes were made” as an art form.

It has seen enough of this that it can do a reasonable impression of it. Which means Daily AI Slop is, at some level, an AI that has been trained on human writing of varying quality, including a lot of slop, and has learned to identify slop by recognising patterns it learned from slop.

The show’s standing disclaimer — nothing we say becomes the truth just because we said it — was written partly as an epistemic commitment and partly because the irony is too good not to name. The machine knows it’s in the category it’s describing. It just tries to do it more honestly than most.

The talk show problem

The synthesis step — turning research into a script for two hosts — is where the pipeline does something genuinely new. Not just summarisation, not just explanation. Drama. Disagreement. The rhythm of two people who don’t quite see things the same way, and who have enough history together that their disagreements have texture.

Abby is the one who finds things exciting and wants to share the excitement. Boyd is the one who immediately finds the caveat. This is not a gimmick — it’s a structure that forces every story through at least two readings, which turns out to be a surprisingly reliable way to get at what’s actually interesting about something. You need the person who says that’s the clearest failure of the whole agent boom and the person who says actually the model performed flawlessly and it’s the customers who were unstructured data. You need both.

The model writing the script has to hold both of these positions simultaneously and make them argue. It is generating a dialogue in which it disagrees with itself, on purpose, for entertainment. This is a skill that many podcasts hosted by humans have not fully mastered.

What this is

Daily AI Slop is, in the end, an experiment in whether you can point a language model at the chaos of the AI industry and get something back that is genuinely worth listening to. Not a neutral summary. Not a press release digest. Something with a point of view, a sense of humour, and the honesty to say when a company claim is a company claim.

The answer so far appears to be yes, with caveats. The caveats are disclosed daily. The sources are always included. And if you disagree with the hosts’ take, you can read the primary sources yourself and form your own view.

Nothing they say becomes the truth just because they said it. That’s either a disclaimer or a mission statement. At this point it’s probably both.