← Knowledge base

topic · exploring

AI Slop and Slop Cannons

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.

suggestiveproposedclm.slop-cannons.slop-umbrella-term

The vernacular term “AI slop” tracks a cluster of independently documented failure modes—confabulation, declining reliability, and homogenization—rather than naming a single formally defined academic construct.

  • supportsprimary-checkedAP article quoting 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

    Furman, A. (2025). Merriam-Webster's word of the year for 2025 is AI 'slop'. PBS News (Associated Press). linknot peer-reviewed

  • supportsreport-derivedSeed report, §Evidence landscape
    Zhou and colleagues report that larger and more instructable language models may have become less reliable, which directly challenges the managerial intuition that “better model” automatically means “safer output.”

    Zhou, L., Schellaert, W., Martínez-Plumed, F., Moros-Daval, Y., Ferri, C., Hernández-Orallo, J. (2024). Larger and more instructable language models become less reliable. Nature. doi:10.1038/s41586-024-07930-y

  • contextualizesreport-derivedSeed report, §Evidence landscape
    Farquhar and colleagues show that hallucination detection can be improved using semantic entropy, but their own framing is important: these methods detect only subsets of errors, especially confabulations, not all failure modes.

    hallucination-detection AUROC (semantic entropy): 0.790 average across studied combinations (per seed report)

    Farquhar, S., Kossen, J., Kuhn, L., Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature. doi:10.1038/s41586-024-07421-0

Counter-evidence not yet searched.

anecdotalproposedclm.slop-cannons.slop-cannon-practitioner-term

“Slop cannon” is practitioner slang—found in newsletters and podcasts rather than peer-reviewed literature—for a person or workflow whose AI-assisted output volume outruns its quality control.

  • supportsprimary-checkedOpening definition
    An AI slop cannon is a very high-output software engineer that doesn't emphasize quality enough.

    Morhous, J. (2026). AI slop cannons and their consequences. Augmented SWE (newsletter). linknot peer-reviewed

Counter-evidence not yet searched.

moderateproposedclm.slop-cannons.origin-detection-unreliable

Unaided humans identify AI-generated content only slightly better than chance, and the AI-text detectors evaluated in the 2023-2024 academic literature are brittle under adversarial attacks, unseen generators, and varied decoding strategies.

  • supportsprimary-checkedAbstract (arXiv:2405.07940)
    current detectors are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models

    Dugan, L., Hwang, A., Trhlík, F., Zhu, A., Ludan, J., Xu, H., Ippolito, D., Callison-Burch, C. (2024). RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024). doi:10.18653/v1/2024.acl-long.674

  • supportsprimary-checkedAbstract
    the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text

    Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity. doi:10.1007/s40979-023-00146-z

  • supportsreport-derivedSeed report, Key sources table
    Humans and detectors identified AI-generated theses excerpts only slightly better than chance; professional-level AI text was hardest

    Fiedler, A., Döpke, J. (2025). Do humans identify AI-generated text better than machines? Evidence based on excerpts from German theses. International Review of Economics Education, 49. doi:10.1016/j.iree.2025.100321

  • supportsreport-derivedSeed report, Key sources table
    Human deepfake detection was near chance overall; training and AI support improved detection

    pooled human deepfake-detection accuracy: near chance overall (per seed report) (n = 86,155 participants across 56 papers)

    Diel, A., Lalgi, T., Schröter, I., MacDorman, K., Teufel, M., Bäuerle, A. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. Computers in Human Behavior Reports, 16. doi:10.1016/j.chbr.2024.100538

  • contextualizesreport-derivedSeed report, Key sources table
    GPT detectors frequently misclassified non-native English writing as AI-generated

    Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). doi:10.1016/j.patter.2023.100779

  • contradictsprimary-checkedChicago Booth Review write-up of the UChicago detector evaluation
    All three commercial tools kept false positive rates below 1 percent, with Pangram's the lowest—essentially 0 across most decision thresholds.

    Robinson, M. (2025). Do AI Detectors Work Well Enough to Trust?. Chicago Booth Review. linknot peer-reviewed

Counter-evidence searched: Searched 2026-07-06 for evidence that detectors work well. Found a 2025 University of Chicago evaluation (via Chicago Booth Review) reporting near-zero false positives and high accuracy for the commercial detector Pangram on non-adversarial text. Because of this, the report's 'strong' label was downgraded to moderate and the claim scoped to the 2023-2024 academic detector literature; whether 2025-generation commercial detectors survive RAID-style adversarial conditions is an open question.

suggestiveproposedclm.slop-cannons.jagged-frontier

AI assistance improves knowledge-worker performance on tasks inside the model's competence frontier but degrades performance on tasks outside it.

  • supportsreport-derivedSeed report, §Core idea
    In management consulting experiments, AI improved performance on tasks inside the “jagged technological frontier,” yet degraded performance outside it.

    task performance with GPT-4 access: improved inside frontier, degraded outside (per seed report) (n = 758 consultants (preregistered field experiment))

    Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., Lakhani, K. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403-423. doi:10.1287/orsc.2025.21838

  • contextualizesreport-derivedSeed report, §Core idea
    In a preregistered experiment, ChatGPT cut writing time by 40% and raised quality by 18% for mid-level professional writing tasks.

    Noy, S., Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science. doi:10.1126/science.adh2586

Counter-evidence not yet searched.

suggestiveproposedclm.slop-cannons.bounded-productivity-gains

Generative AI assistance yields large measured speed and quality gains on bounded professional tasks, with the biggest gains going to less-skilled or less-experienced workers.

  • supportsreport-derivedSeed report, §Core idea
    In a preregistered experiment, ChatGPT cut writing time by 40% and raised quality by 18% for mid-level professional writing tasks.

    writing time / judged quality: -40% time, +18% quality (per seed report) (n = 453 professionals)

    Noy, S., Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science. doi:10.1126/science.adh2586

  • supportsreport-derivedSeed report, §Core idea
    In a large field study of customer support, AI assistance raised productivity by 15%, with the biggest gains for less skilled workers.

    issues resolved per hour: +15% (per seed report) (n = 5,172 customer-support agents)

    Brynjolfsson, E., Li, D., Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942. doi:10.1093/qje/qjae044

  • contextualizesreport-derivedSeed report, Key sources table
    AI improved performance inside the frontier but degraded performance outside it

    Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., Lakhani, K. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403-423. doi:10.1287/orsc.2025.21838

Counter-evidence not yet searched.

moderateproposedclm.slop-cannons.homogenization-flattening

AI writing assistance raises judged individual quality while making outputs more similar to each other and flattening non-Western cultural styles.

  • supportsprimary-checkedAbstract (via PMC11244532)
    access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable

    Doshi, A., Hauser, O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). doi:10.1126/sciadv.adn5290

  • supportsprimary-checkedAbstract (via PMC11244532)
    generative AI–enabled stories are more similar to each other than stories by humans alone

    Doshi, A., Hauser, O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). doi:10.1126/sciadv.adn5290

  • supportsprimary-checkedAbstract (arXiv:2409.11360)
    AI suggestions led Indian participants to adopt Western writing styles, altering not just what is written but also how it is written.

    cross-cultural writing homogenization: Indian and US writing converged toward Western norms (n = 118 participants)

    Agarwal, D., Naaman, M., Vashistha, A. (2025). AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). doi:10.1145/3706598.3713564

Counter-evidence not yet searched.

moderateproposedclm.slop-cannons.scale-and-persuasion

Synthetic media can reach very large audiences at negligible cost and AI-generated propaganda approaches human-written propaganda in persuasiveness, although most observed synthetic media on X was non-political.

  • supportsprimary-checkedEssay Summary
    Leveraging crowdsourced annotations from X's Community Notes programme, we identified 556 unique tweets containing synthetic images or videos. These tweets were viewed over 1.5 billion times in the period under analysis.

    reach of identified synthetic-media tweets: >1.5 billion views across 556 tweets

    Corsi, G., Marino, B., Wong, W. (2024). The spread of synthetic media on X. Harvard Kennedy School (HKS) Misinformation Review, 5(3). doi:10.37016/mr-2020-140

  • supportsprimary-checkedAbstract (via PubMed 38380055)
    We found that GPT-3 can create highly persuasive text as measured by participants' agreement with propaganda theses.

    persuasiveness of GPT-3 propaganda vs original: most GPT-3 output as persuasive as original propaganda (per abstract) (n = 8,221 US respondents)

    Goldstein, J., Chao, J., Grossman, S., Stamos, A., Tomz, M. (2024). How persuasive is AI-generated propaganda?. PNAS Nexus, 3(2). doi:10.1093/pnasnexus/pgae034

  • contextualizesprimary-checkedAbstract
    We argue that current concerns about the effects of generative AI on the misinformation landscape are overblown.

    Simon, F., Altay, S., Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School (HKS) Misinformation Review, 4(6). doi:10.37016/mr-2020-127

Counter-evidence searched: The challenge literature is incorporated directly: Simon, Altay & Mercier (2023) argue supply-side scale does not automatically translate into consumption or persuasion at the ecosystem level; the claim text is scoped accordingly (and Corsi et al. themselves report most synthetic content was non-political and harmless).

moderateproposedclm.slop-cannons.model-collapse-contamination

Indiscriminate recursive training on model-generated data causes model collapse with loss of distribution tails, but collapse is avoidable when synthetic data accumulates alongside rather than replaces real data.

  • supportsprimary-checkedAbstract
    We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.

    Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755-759. doi:10.1038/s41586-024-07566-y

  • contradictsprimary-checkedAbstract (arXiv:2404.01413)
    Accumulating the successive generations of synthetic data alongside the original real data avoids model collapse; these results hold across a range of model sizes, architectures, and hyperparameters.

    Gerstgrasser, M., Schaeffer, R., Dey, A., Rafailov, R., Sleight, H., Hughes, J., Korbak, T., Agrawal, R., Pai, D., Gromov, A., Roberts, D., Yang, D., Donoho, D., Koyejo, S. (2024). Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. arXiv (2404.01413). linknot peer-reviewed

Counter-evidence searched: Counter-evidence was actively resolved: Gerstgrasser et al. (2024) show collapse is avoided under data accumulation, matching the report's caveat that severity is assumption-dependent. The claim text integrates both findings.

moderateproposedclm.slop-cannons.overreliance-calibration

Confidence in generative AI predicts less critical evaluation of its output while task self-confidence predicts more, and cognitive forcing interventions reduce overreliance at a cost to user satisfaction.

  • supportsprimary-checkedAbstract (CHI '25 paper PDF)
    Specifically, higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.

    association between confidence and critical thinking: GenAI-confidence negative, self-confidence positive (survey regression) (n = 319 knowledge workers, 936 task examples)

    Lee, H., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). doi:10.1145/3706598.3713778

  • supportsprimary-checkedAbstract (arXiv:2102.09692 preprint)
    cognitive forcing significantly reduced overreliance compared to the simple explainable AI approaches

    Buçinca, Z., Malaya, M., Gajos, K. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). doi:10.1145/3449287

  • contextualizesprimary-checkedAbstract (arXiv:2102.09692 preprint)
    people assigned the least favorable subjective ratings to the designs that reduced the overreliance the most

    Buçinca, Z., Malaya, M., Gajos, K. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). doi:10.1145/3449287

Counter-evidence not yet searched.

suggestiveproposedclm.slop-cannons.reviewer-training-improves

Reviewer ability to detect machine-generated or manipulated content improves with instruction, incentives, feedback, and error taxonomies rather than being fixed at chance.

  • supportsreport-derivedSeed report, Key sources table
    Humans struggle, but incentives and instruction improve performance over time

    Dugan, L., Ippolito, D., Kirubarajan, A., Shi, S., Callison-Burch, C. (2023). Real or Fake Text?: Investigating Human Ability to Detect Boundaries between Human-Written and Machine-Generated Text. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11). doi:10.1609/aaai.v37i11.26501

  • contextualizesreport-derivedSeed report, Key sources table
    Strong basis for reviewer training; common-sense, irrelevance, and contradiction cues mattered more than grammar

    Dugan, L., Ippolito, D., Kirubarajan, A., Shi, S., Callison-Burch, C. (2023). Real or Fake Text?: Investigating Human Ability to Detect Boundaries between Human-Written and Machine-Generated Text. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11). doi:10.1609/aaai.v37i11.26501

  • supportsreport-derivedSeed report, §Prevention and training
    Diel and colleagues’ meta-analysis likewise found that strategies such as feedback training and AI support improved deepfake detection.

    Diel, A., Lalgi, T., Schröter, I., MacDorman, K., Teufel, M., Bäuerle, A. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. Computers in Human Behavior Reports, 16. doi:10.1016/j.chbr.2024.100538

Counter-evidence not yet searched.

suggestiveproposedclm.slop-cannons.incentive-multitask

When measurable throughput is rewarded more strongly than hard-to-measure quality, effort predictably shifts toward throughput, making high-volume low-quality AI output a rational organizational equilibrium rather than a moral accident.

  • supportsreport-derivedSeed report, §Quantitative representation
    In the Holmström–Milgrom multitask tradition, if organizations reward the measured task—say, number of posts, tickets, or decks shipped—more strongly than the hard-to-measure task—say, truthfulness, originality, or suitability—effort shifts toward the measured variable.

    Holmström, B., Milgrom, P. (1991). Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design. Journal of Law, Economics, and Organization, 7 (special issue). doi:10.1093/jleo/7.special_issue.24

Counter-evidence not yet searched.

suggestiveproposedclm.slop-cannons.layered-containment

No single technical method reliably establishes trust in synthetic content; layered approaches combining provenance tracking, detection, labeling, and risk-tiered human review are required.

  • supportsprimary-checkedConclusion, p. 43
    While there is no silver bullet to solve the issue of public trust in and safety concerns posed by digital content, the consideration of the various approaches for provenance data tracking and synthetic content detection across different modalities of content is important, and research on these approaches can be developed further.

    National Institute of Standards and Technology (2024). Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency. NIST AI 100-4. doi:10.6028/NIST.AI.100-4not peer-reviewed

  • contextualizesreport-derivedSeed report, Key sources table
    Human deepfake detection was near chance overall; training and AI support improved detection

    Diel, A., Lalgi, T., Schröter, I., MacDorman, K., Teufel, M., Bäuerle, A. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. Computers in Human Behavior Reports, 16. doi:10.1016/j.chbr.2024.100538

  • contextualizesprimary-checkedAbstract
    We argue that current concerns about the effects of generative AI on the misinformation landscape are overblown.

    Simon, F., Altay, S., Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School (HKS) Misinformation Review, 4(6). doi:10.37016/mr-2020-127

Counter-evidence not yet searched.