Expert SpeakDigital Frontiers
Published on Nov 06, 2026
As AI-generated content becomes easier to produce and harder to distinguish from human writing, digital platforms face a growing challenge: curbing low-value, synthetic content without treating all uses of AI as harmful
The rapid spread of generative artificial intelligence (AI) has made it possible to produce large volumes of text, images, and other online content at very low cost. However, it has also led to what is increasingly being described as AI slop, which is low-effort, repetitive, or generic content generated at scale and often designed to capture attention rather than contribute meaningful information.
The concern about AI slop is no longer limited to debates about content quality. Several platforms are building systems to identify and reduce such material. LinkedIn offers one recent example. In July 2026, the platform introduced a ‘Seems Like AI Slop’ feedback option and began using member reports alongside classifiers and automated defences to reduce the visibility of low-value content. In August 2026, LinkedIn reported that more than one million members had used this option within the first two weeks and later said that views of content it classified as AI slop had fallen by 40 percent.
Other platforms have addressed the same concern, but without using the term ‘AI slop’. In July 2025, YouTube clarified that its monetisation policy on inauthentic content covers repetitive or mass-produced material. Such content is ineligible for monetisation, while YouTube continues to permit creators to use AI tools. Even Google Search treats the production of large amounts of unoriginal content primarily intended to manipulate rankings as scaled content abuse; its policy expressly lists the use of generative AI to produce many pages without adding value for users as an example of this practice. The policy difficulty is that what constitutes slop does not depend simply on whether AI was involved.
These approaches point to a wider governance question: what features or uses of AI-generated content justify intervention, even as the boundary between recognisably machine-written and human-written content becomes blurred?
AI Use is Not the Benchmark
The emerging platform policies increasingly distinguish between content that serves users and low-value content produced at scale for the purpose of manipulating search results.
YouTube makes this clear, since its inauthentic content rules apply to mass-produced or repetitive material regardless of the method of production, while channels using AI remain eligible for monetisation. Google follows a similar principle—its scaled content policy applies to low-value material, whether produced through automation, human effort, or a combination of both.
This distinction matters because provenance and conduct answer different questions. Knowing that generative AI contributed to a post is relevant to its production process. It does not categorise the post as repetitive, useful, manipulative, or deserving of reduced distribution.
To pinpoint actual slop, patterns such as mass production, repeated posting, automated engagement, or attempts to game visibility and attention should be examined. It also helps separate these practices from non-harmful AI uses, such as using it to draft, edit, or improve a piece of writing.
However, a new problem is emerging. In instances where platforms still want to identify AI-generated material, the stylistic features that make machine-written text recognisable can themselves be learned, removed, and deliberately disguised by AI.
What Makes Writing Look Like AI?
AI text detection often relies on patterns in style rather than any definitive marker of machine authorship. Research has found differences in vocabulary, formality, originality, sentence construction, and other stylistic features between human- and AI-generated writing. A 2025 ACL study involving 300 non-fiction articles found that frequent users of generative AI were particularly good at identifying machine-written text. Their judgements drew not only on recurring AI vocabulary, but also on broader characteristics such as formality, clarity, and originality.
Yet, these markers are not fully accurate. Another 2025 study found that the accuracy of AI-text detectors is varied, depending on the writing style and text complexity; in some settings, their results seemed akin to random categorisation. The research also found that easier-to-read writing could be disproportionately misclassified as AI-generated.
This exposes a basic weakness in using stylistic appearance as a determinant of authorship. The features that make text appear machine-written can vary across domains, models, and writing styles. More importantly, once those features are understood, they can also be deliberately changed.
The Humanisation Arms Race
The problem is no longer limited to the fact that AI detectors can make mistakes. AI-generated text can now be produced in ways that make detection significantly harder and rewritten in a manner that emulates natural human writing, known as humanisation of AI.
In September 2026, Bhuwan Dhingra of Duke University and Danish Pruthi of the Indian Institute of Science published research examining whether an AI system could make its output less detectable by relying heavily on text generated by a separate base language model. They tested this approach across 400 tasks. Pangram, an AI detector, flagged 77 percent of the answers Claude wrote on its own, while flagging only 24 percent of the answers Claude put together using text from other AI models. Three other detectors identified fewer than 10 percent. The paper has not yet been peer-reviewed.
The findings matter as they show that technology has moved beyond the familiar practice of rewriting AI-generated text after it has been produced. Here, the AI system drew heavily on text from a base model while generating the response itself, with up to 90 percent of the final output coming from that second model. The generated text remained relatively effective across several tasks while becoming substantially harder for detectors to identify.
If AI-generated content can be made less recognisable at the point of generation, whether it appears machine-written becomes a less reliable basis for determining how it should be treated.
This points to a deeper challenge for AI detection. Systems learn to recognise patterns associated with machine-written text, but new-generation techniques can be designed to avoid those same patterns. For AI-slop governance, that makes stylistic detection increasingly unreliable.If AI-generated content can be made less recognisable at the point of generation, whether it appears machine-written becomes a less reliable basis for determining how it should be treated.
Governments Are Regulating Synthetic Content, not ‘Slop’
Many governments worldwide have responded to the growing difficulty of distinguishing synthetic from human-created content, although their rules generally focus on identification and transparency rather than content quality.
China has adopted one of the broadest approaches. Its “Measures for Labelling AI-Generated and Synthetic Content,” in force since September 2025, cover AI-generated text, images, audio, video, and virtual scenes. Providers must use visible labels in specific circumstances and embed information in metadata, while users publishing AI-generated material are required to declare it. The rules also prohibit the malicious removal, alteration, or concealment of required labels.
The European Union (EU) has taken a similar provenance-based approach. Article 50 of the AI Act, applicable from August 2026, requires providers of systems generating synthetic text, audio, images, or video to make their outputs machine-readable and detectable as artificially generated or manipulated. It also requires disclosure for certain AI-generated texts published to inform the public on matters of public interest, subject to exceptions including human review and editorial responsibility.
India’s framework shows that rules focused on proving content source—whether it is AI-generated, altered, watermarked, or labelled—can only be partially effective. Its 2025 AI Governance Guidelines discuss tools such as watermarking, provenance markers, and unique identifiers, while also recognising that these safeguards can be bypassed. The 2026 amendments to the IT Rules have introduced more concrete obligations for synthetically generated information, but their scope is limited to audio, visual, and audiovisual content. AI-generated text sits outside their purview.
While these approaches address an important problem—helping users know when content is synthetic—they do not determine whether the content is repetitive, low-value, or produced at a scale that degrades an information environment.
From Identifying AI to Governing Its Use
Humanisation shows why transparency and detection cannot be the only responses. Machine-readable markers and metadata can provide stronger evidence of provenance than writing style alone, and rules such as those in China and the EU attempt to preserve such signals. Where platforms rely on language to infer AI authorship, increasingly sophisticated generation and rewriting techniques make assessment less dependable.
Even perfect identification of AI-generated content would not resolve the core issue of AI slop. A label can tell a user how the content was produced; it cannot identify whether the material is informative, repetitive, manipulative, or one of hundreds of near-identical automated posts.
It may therefore no longer be enough to ask whether content was AI-produced. As AI-generated text becomes harder to distinguish from human writing, platforms will also need to look at how that content is shared and used at scale. The solution does not lie in policing every piece of content, but in identifying when automated production begins to undermine the quality of online spaces.
Tanusha Tyagiis a Research Assistant with the Centre for Digital Societies at the Observer Research Foundation.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
Author
Tanusha Tyagi
Tanusha Tyagi is a research assistant with the Centre for Digital Societies at ORF. Her research focuses on issues of emerging technologies, data protection and …
Related Search Terms
Publications
Start at the Source: Why India’s Waste-to-Energy Push Needs Segregation First
Sustainable Development
Nov 08, 2026
