# The AI Slop Report 2026 AI slop is an unpriced transfer of work. Authorship is the investment that prevents it. - Canonical: https://paciva.ai/resources/ai-slop-2026/ - Publisher: Paciva (paciva.ai). Report ASR-2026-09, published September 15, 2026. Sources rechecked September 11, 2026. - This is the machine readable companion to the interactive report at the canonical URL. ## The claim structure The memorable line: slop is work whose producer has not invested in making it useful. Investment means judgment, context, evidence, testing, revision, acceptance rules, monitoring, or accountability; it is not measured in keystrokes or elapsed time, and the producer is the accountable source, a person or an institution, never the language model. The operational definition: slop is a failed handoff that imposes avoidable interpretation, verification, repair, delay, or trust costs because the work was not made useful enough for its recipient, purpose, and level of risk. AI slop is that failure with the work materially generated, transformed, or scaled by AI. AI provenance alone is never sufficient; the handoff must also fail the recipient burden test. The proposed category: the recipient tax, the portion of a recipient's burden that a proportionate, producer owned upstream control could likely have reduced at lower total system cost. Legitimate judgment, joint uncertainty, and inherent complexity are ordinary coordination, not tax. The report proposes two sided protection: prevent low investment work before handoff, and protect people from imported work across their connected surfaces. The governing measure: realized useful outcome net of producer, reviewer, defense, recipient, defect, and delay costs, reported with its distribution across senders and recipients. ## Section summaries 1. The sender saved minutes. The recipient got the invoice. In a commercial survey of 1,150 full time US desk workers (BetterUp Labs and Stanford Social Media Lab, 2025), 40 percent reported receiving workslop in the prior month, with nearly two hours self estimated per incident and a modeled $186 per month per affected employee. Self reports and a vendor cost model, not audits. The recipient's work: detect, interpret, reconstruct, verify, repair, respond, update trust. Historical analogy: 2012 spam economics found roughly $20 billion in annual US costs against $200 million in spammer revenue, about 100 to 1 (Rao and Reiley, JEP). The interactive calculator reproduces the $186 model at its defaults. 2. The flood is measurable. Slop is not one number. Detector estimates: AI attributed posts reached 37.03 percent on Medium and 38.95 percent on Quora by October 2024 (ACL 2025); roughly half of new articles primarily AI generated by 2025, then plateauing (Graphite). NewsGuard's discovered AI content farm tally grew from 49 sites in April 2023 to 3,749 on June 23, 2026. A manual audit found 5.3 percent of 1,082 biomedical education videos both AI generated and low quality (JMIR). Every source sits on four separate axes: provenance, quality, deceptive intent, recipient burden. No study measures all four. No universal slop rate is supported and the report states none. 3. Slop begins at the handoff. Private drafts tax nobody. A recipient centered taxonomy: context, epistemic, relationship, judgment, volume, execution, and ecosystem slop, each named by the work the recipient inherits. One illustrative brief sent three ways: human slop, AI slop, authored AI assistance. Bad output alone does not prove a lazy producer; ask who controlled the deadline, workload, and release decision (de Fine Licht, Philosophy and Technology, 2026). Wanted synthetic entertainment is the negative control. 4. Why rational systems produce irrational amounts of slop. One journal saw a 42 percent submission increase after ChatGPT with detector banded desk rejections (Organization Science, 2026). Consultants with AI finished 12.2 percent more tasks and worked 25.1 percent faster inside the tested frontier, and were 19 percentage points less correct on one task outside it (Dell'Acqua et al.). In a randomized 680 person study, AI interventions raised producer side signals while none improved recipient perceptions and all increased dislikes (Moller et al., Scientific Reports, 2026). An incentive design problem before a writing problem. 5. The recipient tax compounds. Four levels: individual, relational, institutional, ecosystem. Recursive training on generated data degrades information unless human data is retained (Shumailov et al., Nature, 2024); external verification can change the outcome, with limits from imperfect verifiers (Yi et al., ICLR 2026). Adverse selection is named as plausible, not established. Measure the total and its distribution, not the average. 6. Authorship is the control system. Adapted from ICMJE: substantive contribution, critical review, final approval, accountability for repair. Three states: individual authorship, institutional authorship, pseudo authorship. Passive review fails: wrong AI answers were caught 8 percent of the time with ordinary explanations, 27 percent with cognitive forcing, 49 percent without AI (Bucinca et al.). AI primary drafting reduced ownership and accountability more than editing assistance (Li et al., CHI 2024). Accessibility is part of authorship (Zhao et al.). Detection is not authorship: current detector evaluations show inconsistent biases (Stowe et al., ACL 2026) and a Czech evaluation found no systematic bias (Al Ali et al.). The villain is abandoned responsibility, not saved time. 7. Four phases, ten gates. Intend (outcome, recipient), Challenge (ideation, stress test), Prove (plan, acceptance, match effort, assess), Own (accountable release, abstain). A proposed synthesis extending adjacent evidence (Bucinca; Tessler et al., Science 2024; Tas, Memmert, and Bittner, Electronic Markets 2026). It has not been shown to reduce recipient tax. Match effort and abstain keep the method from becoming process slop. 8. Two-sided protection. Recipients include queues, institutions, and information systems, not only people. Sender side prevention plus recipient side defense, meeting at the handoff. No product performance claim; protection earns its place only when burden removed exceeds producer, reviewer, and defense burden added. 9. Measure useful outcomes, not generated output. Governing metric: total system burden per realized useful outcome, with distribution. Denominator guardrails: predefine eligible handoffs and the window; count eligible, useful, unsuccessful, unresolved, and excluded with reasons; keep failed work in total burden. Five minimum measures: time to useful outcome, recipient rework, escaped defects and repair, producer and review overhead, defense errors. Bounded sequence: define, baseline, apply controls, compare, then expand, revise, or stop. 10. Abundance changes what is scarce. Production is abundant; judgment, context, and responsibility are scarce. One reader action: nominate one costly handoff for a bounded baseline and comparison. ## Questions readers ask - What is AI slop? A failed handoff whose work was materially generated, transformed, or scaled by AI and which imposes avoidable interpretation, verification, repair, delay, or trust costs. Provenance alone is not enough. - What is the recipient tax? The portion of recipient burden a proportionate, producer owned upstream control could likely have reduced at lower total system cost. - Is AI generated content always slop? No. Human work can be slop, AI assisted work can retain authorship, and wanted synthetic entertainment is not slop. - How much does workslop cost? The measured signal: 40 percent of 1,150 surveyed US desk workers, nearly two hours per incident, $186 modeled per affected employee per month. The total economic tax is unmeasured. - Can AI detectors identify slop? No. Detectors estimate provenance, not usefulness, with inconsistent biases across systems and populations. - What is the fastest way to act? Nominate one handoff, run the bounded diagnostic, then expand, revise, or stop. ## Evidence classes 28 ledgered claims: peer reviewed studies, commercial measurements labeled as such, accepted conference papers, authoritative standards, and report synthesis labeled as synthesis. Detector estimates are never converted into slop prevalence. The specimens on the report cover are recreated genre examples built for this report; none reproduces a real post, article, or message. ## About Paciva Paciva makes Pax, an executive assistant for individuals and a chief of staff to organizations. The report argues the doctrine generically and makes no product performance claim. Contact: https://paciva.ai/ and https://calendly.com/jeremy-paciva/business-review-conversation