The era of artificial intelligence slop: what makes content worth publishing

Published 19 August 20266 min read

Written by

Fahad IshtiaqFounder, Vela

A business can now produce more articles, messages and reports with less drafting effort. Whether anyone benefits from the additional output is a separate question.

The word “slop” captures a particular failure: material produced with little care for its usefulness or the effort it imposes on others. Merriam-Webster selected it as its 2025 word of the year, while Macquarie chose “AI slop” (Furman, 2025; Macquarie Dictionary, 2025). The term names a frustration readers already recognise: more material to get through, with little reason to keep reading.

Simon Willison’s early discussion emphasised unwanted, unreviewed AI-generated material (Willison, 2024). Coverage in The Guardian and The New York Times helped bring the term to a wider audience (Hern & Milmo, 2024; Hoffman, 2024).

The distinction matters for businesses using AI. Generated text can be useful, and human-written text can be empty. The editorial question is whether a piece gives its intended reader something accurate, relevant and worth the time required to understand it.

Production at a new scale

Attempts to measure AI-generated content illustrate the scale of production and the limits of detection.

Graphite’s analysis of English-language articles sampled from Common Crawl estimates that around half of newly published articles were primarily AI-generated by early 2026. It uses three detectors and reports tests of their accuracy (Paredes et al., 2026). The estimate classifies authorship in that sample; it does not measure quality.

NewsGuard separately reported identifying more than 3,000 AI content-farm sites by March 2026 (NewsGuard, 2026). The sites meet NewsGuard’s criteria for low-quality, AI-generated publishing.

Music offers another view of supply. Deezer reported that fully AI-generated tracks exceeded half of daily new deliveries on peak days in June 2026, with roughly 90,000 such tracks arriving per day (Deezer, 2026; Mehta, 2026). Spotify reported removing more than 75 million spam tracks over twelve months (Spotify, 2025). Spam is the relevant category in Spotify’s figure; it should not be read as a count of confirmed AI-generated tracks.

Academic writing is also changing. Research analysing more than 1.1 million papers estimated growing use of LLM-modified language (Liang et al., 2025). A separate study estimated that 6.5% to 16.9% of text in reviews at four AI conferences may have been substantially modified by LLMs (Liang et al., 2024). These studies measure changes in language, leaving the quality of the underlying research and review to be assessed separately.

Across publishing, music and research, production is changing quickly. The editorial task is to decide which of that output deserves an audience.

The effort can move to the reader

For a business, the most immediate cost may be the work an output creates downstream.

BetterUp and Stanford researchers use “workslop” for professional material that looks complete but lacks the substance needed to advance a task. Their survey-based research describes recipients spending effort interpreting and repairing such work, alongside negative judgements about its senders (BetterUp, 2025; Niederhoffer et al., 2025).

Consider a hypothetical market-analysis report. It arrives on time, uses polished language and recommends three priorities. Yet its figures lack traceable sources, its categories overlap and its recommendations could apply to almost any company. The author has finished drafting; the recipient must now perform the analysis.

That is the operational failure worth addressing. The relevant measure is how much useful work the document enables after verification, interpretation and correction are counted. A longer report may simply provide more surface area for the missing reasoning.

Abundance does not establish demand

Producing a large share of new material does not mean attracting a corresponding share of attention.

In a separate study, Graphite classified 86% of sampled articles ranking in Google Search as human-written and 14% as AI-generated. The study did not evaluate heavily human-edited AI assistance (Graphite, 2025). In that sample, human-written articles still occupied most ranking positions.

Deezer likewise reports a substantial difference between AI-generated uploads and their share of listening, alongside extensive fraudulent streaming among those tracks (Deezer, 2026). On that platform, a growing supply has yet to attract a comparable share of listening.

Herbert Simon’s argument about the relationship between abundant information and scarce attention remains useful here (Simon, 1971). More supply creates a stronger practical reason to consider what deserves a reader’s attention.

For a business publishing an article, the useful question is what a reader can understand or decide after reading it. An editorial review should assess that benefit separately from publication frequency or search performance.

Filters, provenance and labels do different jobs

Platforms have responded to low-value production. Google introduced rules against scaled content abuse and reported reducing low-quality, unoriginal search content by 45% after its 2024 changes (Tucker, 2024). The figure describes Google’s assessment of those search changes.

YouTube’s monetisation policy addresses mass-produced or repetitive content, and its 2026 leadership letter discusses reducing low-quality AI material while also supporting AI creation tools (Mohan, 2026; YouTube, 2025). These policies distinguish harmful publishing patterns from the mere use of a tool.

Provenance addresses another problem. C2PA Content Credentials can record aspects of a file’s origin and editing history. The standard does not determine whether its claims are true (Coalition for Content Provenance and Authenticity, n.d.). A verifiable history helps answer where material came from; the content still needs assessment.

Labels also influence judgement. In experiments involving 4,976 US and UK participants, Altay and Gilardi found that an AI-generated label reduced perceived headline accuracy even when a headline was true or human-made. Participants tended to interpret the label as implying full automation (Altay & Gilardi, 2024).

Poetry research adds a related complication: non-expert readers often preferred the tested AI poems, yet ratings were affected by what readers were told about authorship (Porter & Machery, 2024). Preferences about origin and assessments of quality are related but distinct.

Public concern is real. Pew found that 76% of Americans considered it very or extremely important to distinguish AI-made material from human-made material, while 53% lacked confidence in doing so (Kennedy et al., 2025). The Reuters Institute’s 2025 report also documents concern about distinguishing true and false online information and expectations that AI-produced news would be less trustworthy (Newman et al., 2025). Readers want to know where material comes from and whether they can trust what it says.

What a publishing standard should protect

Technical research supplies a further reason to preserve quality. Shumailov and colleagues demonstrated deterioration under recursive training on model-generated data (Shumailov et al., 2024). Gerstgrasser and colleagues showed that accumulating synthetic data alongside retained real data could avoid collapse in their studied settings (Gerstgrasser et al., 2024). The treatment of source data matters to the quality of what a model learns.

For publishers, a more immediate responsibility is to maintain the connection between claims and evidence. Before releasing an article, establish its central argument, check consequential facts, examine whether its sources support its wording, and remove passages that add length without advancing understanding.

A named author identifies who accepts responsibility, and a reference list gives readers a route to the evidence. Their value depends on the accuracy of the claims and the care taken in reviewing them. Sources can be misread, and confident writing can conceal weak reasoning regardless of how the first draft was made.

AI can assist with research organisation, drafting and editing while those standards remain in force. Faster drafting gives an editor more room to question the argument, test the evidence and make the explanation worth reading.

A business should publish when it has something worth explaining and can stand behind the explanation. The first draft may be cheap. The reader’s attention still has to be earned.

Sources

Every figure in this article was checked against these sources.

All posts