AI slop and system overload

This thread follows a side effect of cheap generation: not any single fake, but the sheer volume of text, images, audio and code that models can now produce, arriving at systems built for human rates of output. The problem is distinct from the deepfakes and misuse collected elsewhere — the content is often not even malicious, merely low-value and endless — and it showed up first wherever a queue depended on human attention. Within a week of ChatGPT’s launch, Stack Overflow banned model-generated answers because plausible wrong ones arrived faster than moderators could check them; months later the science-fiction magazine Clarkesworld closed submissions after machine-written stories overran its slush pile.

Institutions built around a fixed human throughput met the same wave. Courts absorbed fabricated case law from both represented parties — the lawyers sanctioned in Mata v. Avianca — and self-represented ones, as when a UK tax tribunal found an appellant had cited nine judgments that did not exist. Scientific publishing hit its own limit when Wiley closed nineteen journals after paper mills using AI overwhelmed peer review; Google rewrote its spam rules to police content produced at scale regardless of whether a human or a machine made it; and Wikipedia’s editors codified a fast-track deletion rule for obviously machine-written articles. The response was almost always the same shape — a ban, a queue closure, a new screening policy — because the receiving system could not simply scale to match.

By 2025 the phenomenon had a name: Merriam-Webster made “slop” its word of the year. It had also acquired measurable scale. A Stanford study put a price on “workslop” circulating inside companies; the maintainer of the networking tool curl shut down a seven-year bug bounty after machine-generated vulnerability reports made triage unsustainable; and on music platforms the share of AI-generated uploads climbed until Deezer reported it had passed half of everything submitted daily. The throughline is a single shift in what was scarce: once producing passable content stopped being the constraint, the constraint became whether the human systems meant to read, judge, host and curate it could still keep up.