AI Slop and Slop Cannons
What we actually think: "AI slop" is vernacular, not a formal construct — Merriam-Webster's
2025 word-of-the-year definition ("digital content of low quality that is produced usually in
quantity by means of artificial intelligence") is the closest thing to a canonical anchor — but
the phenomena it bundles are each independently documented: models can be confidently wrong and
may get less reliable as they get more instructable (Zhou et al. 2024), hallucination detection
catches only subsets of errors (Farquhar et al. 2024), and AI assistance measurably homogenizes
output, individually better but collectively more similar (Doshi & Hauser 2024) and flattened
toward Western styles (Agarwal, Naaman & Vashistha 2025). The most defensible definition is an
outcome category, not an authorship category: output whose apparent polish exceeds its epistemic,
functional, or contextual quality.
A "slop cannon" — the essay's central figure — is strictly practitioner slang; the earliest
verifiable usages we found are newsletters and podcasts (e.g., Morhous 2026), and the phrase has
no peer-reviewed standing. But the pattern it names is well grounded: generation costs collapsed
(large measured speed gains in writing and support work) while review remained effortful and
confidence-sensitive (Lee et al. 2025; Buçinca et al. 2021). A slop cannon is what you get when
an organization pushes volume through that asymmetry. The incentive logic (Holmström–Milgrom
multitask theory) says this is an equilibrium, not an accident — though for now that leg of the
argument is theoretical, and we hold it at "suggestive."
The single most important operational finding is that detecting slop is not the same as
detecting AI use. Humans identify AI text and deepfakes near chance (Fiedler & Döpke 2025; Diel
et al. 2024), and the 2023–2024 academic detector literature (RAID; Weber-Wulff et al.) found
detectors brittle and biased against non-native writers (Liang et al. 2023). Our counter-evidence
search complicated this, however: a 2025 University of Chicago evaluation reports near-zero false
positives for the commercial detector Pangram on non-adversarial text. We therefore hold the
detection claim at "moderate," scoped to the earlier detector generation, and treat defect-first
review (grounding, contradiction, task fit) as the robust conclusion either way — it does not
depend on which detector generation wins.
On systemic risk, the evidence supports concern without catastrophism. Synthetic media achieved
1.5B+ views from just 556 tweets on X (Corsi et al. 2024) and GPT-3 propaganda was nearly as
persuasive as the human-written original (Goldstein et al. 2024) — but most observed synthetic
content was non-political, and Simon, Altay & Mercier (2023) argue the misinformation fears are
overblown at the ecosystem level. Recursive-training collapse is real under indiscriminate
replacement of real data (Shumailov et al. 2024) but avoidable under data accumulation
(Gerstgrasser et al. 2024). The report's hazmat metaphor survives as an operational stance —
containment, labeling, provenance, risk-tiered review, per NIST AI 100-4's "no silver bullet"
conclusion — but not as an ontological claim that AI output is inherently toxic.
Limitations of this ingestion: the seed report's citations were opaque citeturn tokens, so every
bibliographic entry was rebuilt independently; two rows (the "slop cannon" phrase sources and the
X synthetic-media study) were resolved to real publications by search rather than by following
the report's links. The three flagship economics papers (Noy & Zhang; Brynjolfsson, Li & Raymond;
Dell'Acqua et al.) could not be primary-checked because every accessible copy is paywalled or
bot-blocked, so the two productivity claims sit at "suggestive" despite the report labeling them
"strong" — that is a verification limit, not a judgment that the findings are weak, and upgrading
them is the highest-priority next research action. The report's quantitative models (slop score,
expected-loss equation, review decision rule, signal-detection framing) are managerial syntheses
the report itself marks as theoretical; we did not encode them as claims.