The Death of Slop: Why Detection Loses and Provenance Might Not

The Flood Arrived Exactly On Schedule

Open any large social feed now and the pattern is hard to miss. Listicles with no authorial fingerprint. Product round-ups apparently written by nobody in particular. AI-narrated clips stitched from stock footage. News summaries that repeat a wire report badly, then acquire a few invented details on the journey. The internet has not merely become noisier; it has become mechanically prolific.

That outcome is often framed as a technological mishap, as if the models escaped the lab and left everyone scrambling. In truth it is more mundane, and more structural. The web was built on an attention market. If attention can be bought, sold, measured and routed into advertising or subscriptions, then any fall in the cost of producing attention-capturing artefacts will eventually be exploited. Generative AI did not create that logic. It simply removed the last meaningful production constraint.

This is why the current debate so often feels unsatisfying. Too much of it assumes the problem is that machines can now write, draw, narrate and remix. The real problem is that generation became cheap while distribution remained effectively free and monetisable. Slop is an economics problem wearing a technology costume.

That distinction matters because it tells you what will fail. If the incentive is to fill the pipes with low-cost material in the hope that a small percentage captures clicks, then any response centred on identifying and suppressing AI output as a category is already aimed at the wrong target. Detection can trim the edges. It cannot reverse the underlying return on investment.

The Cost Curve Moved; The Incentive Did Not

For most of the commercial web, there was a natural throttle on volume. Producing content at scale required labour: reporters, copywriters, editors, designers, videographers, subject specialists. Even mediocre content had a cost floor. That floor imposed discipline. Not good discipline, necessarily, but enough to stop infinite output.

Generative systems collapsed that floor. A single operator can now produce hundreds of product descriptions, location pages, SEO posts, ad variants or social clips in the time it once took to draft a handful. The point is not that all of this content is worthless. Some of it is efficient, useful and entirely defensible. The point is that the marginal cost of trying has fallen towards zero while the upside from one item going viral, ranking in search, or earning affiliate revenue remains intact.

The evidence is all around the market. Search results for commercial queries have become crowded with pages whose purpose is not to inform but to intercept. Social platforms are full of engagement bait generated at industrial scale because a tiny conversion rate is enough when generation and publishing cost almost nothing. Amazon has had to contend with AI-generated books and low-value listings. Teachers, researchers and moderators report torrents of machine-produced submissions, comments and spam. OpenAI, Google, Anthropic and others have made creation easier; the business models of platforms made overproduction rational.

That is why slop is not confined to one medium. It appears in text, image, audio and video because the pattern is economic, not stylistic. Wherever low-cost generation meets algorithmic distribution and some form of payment for attention, slop follows.

Why Detection Loses, Even When It Improves

The instinctive response is still to build a better detector: spot the generated artefact, label it, downgrade it, perhaps remove it. This is appealing because it resembles a technical solution to a technical problem. It is also, in the strong sense, unwinnable.

The first reason is temporal. Classifiers learn from known outputs; generators keep changing. That asymmetry is permanent. It is not a bug in present-day systems that better engineering will eliminate. Detection is trained on yesterday’s artefacts and deployed into tomorrow’s distribution environment. Watermarks may help in controlled ecosystems, but on the open web content is copied, cropped, transcribed, translated, reformatted and re-exported. Even robust signals degrade once content leaves the system that created it.

The second reason is conceptual. “AI-generated” is simply the wrong category if the thing one cares about is value. Plenty of machine-assisted content is useful: translated documentation, accessible summaries, synthetic voiceovers for public-service information, draft internal reports checked by experts. Equally, plenty of wholly human content is trivial, manipulative or false. A system that sorts on mode of production rather than reliability, originality, evidence or accountability will make both kinds of mistake: it will suppress useful material and let rubbish through. The more platforms punish “AI-ness” as such, the more they encourage concealment rather than quality.

You do not solve slop by identifying AI perfectly; you solve it by making low-trust distribution less profitable.

The third reason is adversarial economics. Anyone profiting from the flood has strong incentives to evade the filter and only has to succeed item by item. The defender must maintain a broadly accurate system across millions or billions of pieces of content. This is the same asymmetry that shaped spam, fraud and search manipulation. Detection becomes a cost centre for the platform; evasion remains a profit centre for the publisher.

Recent experience bears this out. AI text detectors have repeatedly produced false positives on non-native English writing, formulaic educational prose and plain bureaucratic language. Several universities and schools that flirted with automated detection have discovered the governance problem the hard way: a probabilistic signal is not a fair basis for high-stakes sanction. Even platform labelling is uneven. YouTube asks creators to disclose “realistic altered or synthetic content” in certain circumstances, while TikTok and Meta have introduced synthetic-media labels and policies, but none of these systems comes close to comprehensively identifying low-value generated material, nor could they. They are governance tools at the margin, not a durable answer to abundance.

Provenance Is Not Truth, But It Is Something Machines Can Check

If detection asks the wrong question, what is the right one? Not “was this generated?” but “who is willing to stand behind it, and what exactly are they claiming?” That is the provenance turn.

Provenance is attractive because it shifts the problem from inference to attestation. Instead of guessing from surface features whether a piece of content emerged from a model, a provenance system records claims about origin, modification and review in a form that can be verified. The key move is not mystical. It is administrative. Someone asserts: this file came from this organisation; these edits were made; these sources were consulted; this human reviewed these parts; this version supersedes that one. Those assertions can be signed, time-stamped and checked later.

There are already meaningful building blocks. The C2PA specification, backed by a coalition including Adobe, Microsoft, the BBC, Intel, Arm, Nikon and others, provides a standard for attaching cryptographically verifiable content credentials to media. Adobe’s Content Credentials initiative aims to surface information about how an image or asset was created and edited. Camera manufacturers including Leica and Nikon have announced work that supports provenance workflows for captured images. At the platform layer, provenance features are beginning to appear in creation and editing tools, however unevenly.

This does not prove a claim is true. That is the crucial limitation, and it should be stated without euphemism. Provenance is not epistemology. A signed falsehood is still false. But attribution changes the incentives in a way anonymous generation does not. Unsigned sludge can be sprayed across the network with no reputational balance sheet attached. Signed sludge creates an accountable party, a public record and a potential cost to future distribution.

That is the beginning of a market for trust rather than mere volume. Readers, publishers, search engines, procurement teams and curation agents can distinguish between “unknown origin”, “known origin”, and “known origin with specific review claims”. The content may still be poor; the difference is that it is no longer frictionlessly fungible.

Regulation Is Groping Towards Disclosure, Not Quality

Public policy, notably in Europe, is inching in the same direction. The EU AI Act does not solve slop, nor does it pretend to. But its transparency provisions matter because they normalise the idea that synthetic or manipulated content should in some contexts be disclosed, especially where there is a risk of deception. Separately, the Digital Services Act pushes very large platforms towards greater accountability for systemic risks, algorithmic processes and content governance.

These measures are often discussed as speech regulation. In practice, they are also market-shaping mechanisms. Once disclosure and traceability become ordinary compliance expectations, provenance infrastructure becomes more valuable. So do systems that can ingest provenance claims and act on them.

One should not overstate the effect. Regulation can require labels, reporting and process; it cannot legislate discernment into existence. Nor can it easily police millions of low-value websites, cross-border spam operations or synthetic media passed through a dozen reposting layers. But policy can make the absence of provenance more salient, especially for professional publishing, news, public-sector communications, finance, health and education. In those domains, checkable attribution has practical value even before it has universal adoption.

That is enough to matter. The route out of slop does not require every meme, blog post and affiliate page on the web to become neatly credentialled. It requires enough high-value contexts to reward verifiable origin, and enough downstream tools to preferentially surface it.

The Real Inversion: Move Curation To The Reader’s Side

Provenance is not truth, but it is a form of accountability that machines can check and markets can price.

Provenance addresses attribution. It does not by itself solve distribution. For that, the more radical shift is moving curation away from the platform and towards the reader.

This is easy to describe and harder to absorb because it reverses a long-settled assumption. Today, most people receive information through feeds ranked by entities whose revenues increase with engagement, session length, advertising yield or some closely related proxy. Those firms are not evil for doing so; they are behaving according to their incentives. But they are structurally misaligned with the user who wants relevance, accuracy, variety and calm rather than maximum stimulation.

A reader-side agent changes the principal-agent relationship. Instead of a platform deciding what billions should probably see, an agent acts for one person or one organisation according to explicit instructions. It can be told to prioritise primary sources, to insist on provenance for factual claims, to down-rank anonymous rewrites, to surface opposing views from pre-specified outlets, to exclude whole classes of content farming, or to maintain source diversity targets. Crucially, it can do so consistently.

This is where personal sovereignty in information starts to become operational rather than philosophical. The user is not merely choosing among feeds; they are defining the rules of intake. In Society OS terms, the broad ambition of The Sovereign Standard is to make digital systems answerable to the individual rather than the other way round. In information markets, reader-side curation is one practical expression of that logic.

There are precursors already. RSS was an early, partial form of reader control. Email newsletters and paid subscriptions moved some audiences away from the feed. Browser extensions, read-later services, academic literature tools and enterprise threat-intelligence platforms all offer fragments of user-defined filtering. More recently, AI assistants can summarise, compare and route information on the user’s behalf. What is missing is not capability so much as alignment and governance: the agent must answer to the reader’s intent, not to a hidden optimisation target set by the distributor.

Attack The Economics, Not The Existence, Of Slop

The strongest argument for reader-side curation is economic. Slop does not need to persuade everyone. It only needs cheap generation and indiscriminate access to enough attention to produce a return. If every user had a competent filtering layer that rejected low-trust material, demanded provenance in defined contexts, and learned their standards for evidence, then mass distribution would lose its current advantage.

A flood that reaches nobody useful has no business model.

This is the same reason spam became manageable not when unwanted messages stopped being sent, but when enough of them stopped being delivered, seen or acted upon. Email was not purified. The economics shifted. Sending remained cheap; conversion collapsed.

The analogy is not exact. Slop is broader than spam and often less obviously malicious. Some of it is simply indifferent filler generated because the system rewards volume. But the strategic point holds. One does not have to prevent the production of low-value content to make it unprofitable. One has to reduce the expected yield from blasting it at the public.

That is what provenance plus reader-side curation can do together. Provenance creates machine-readable accountability signals. Reader-side agents use those signals, alongside user preferences and source histories, to decide what crosses the threshold of attention. The result is not a clean internet. It is a more selective market for attention.

The Failure Modes Are Real, And They Matter

This argument should not be romanticised. Reader-side curation and provenance both introduce new risks.

The first is enclosure by preference. A personal agent can build a pleasant intellectual cage if instructed badly or left to optimise for comfort. An information environment that only confirms prior beliefs is not sovereign; it is merely customised. Serious curation tools therefore need explicit pluralism settings: include credible disagreement, inject primary material, reveal what has been excluded, and permit auditing of ranking criteria. Without that, personal filtering becomes a more polite version of algorithmic narrowing.

A flood that reaches nobody useful has no business model.

The second is gatekeeping by attestors. If provenance becomes valuable, whoever controls recognised registries, credential systems or trust lists acquires power. Standards bodies can ossify. Commercial intermediaries can become chokepoints. Smaller publishers, dissidents, freelancers and people outside well-resourced institutions may struggle to obtain credentials that downstream systems favour. The history of identity and payments online offers ample warning: infrastructure built for trust can easily become infrastructure for exclusion.

The third is uneven adoption. High-trust sectors may move first, while the wider web remains saturated. Newsrooms, public agencies, regulated industries and professional creators have stronger reasons to adopt provenance and curation standards. Casual publishing does not. That means the benefits may appear first for those already best served, rather than for the average user in the most polluted parts of the internet.

The fourth is strategic mimicry. Once provenance carries reputational weight, bad actors will counterfeit its social meaning even where they cannot counterfeit its cryptography. Expect badges, pseudo-credentials, friendly-sounding “editorial standards”, and disclosure pages designed to satisfy shallow checks. Human users are susceptible to this; automated agents will be too unless their trust models are carefully designed.

What A More Serious Information Stack Would Look Like

If one were designing for the next phase rather than patching the current one, the components are not especially mysterious.

Creation tools would emit standardised provenance by default where appropriate. Publishers would attach signed assertions not just about authorship, but about review status, source classes and revision history. Browsers, readers and AI agents would expose those signals clearly rather than burying them in menus. Users would be able to set their own policies: always prefer direct documents to summaries; demand attestation for health and financial claims; show me opposing coverage from trusted outlets; reveal when an item lacks provenance.

On the governance side, agent permissions would need to be explicit and revocable. If a curation agent is acting for a user, it should do so under clear authority, with a defined scope, restricted data access, auditable behaviour and the ability to revoke its powers cleanly. In Society OS language, that is where a neutral governance layer such as F-ACT becomes relevant: not as branding, but as a practical way to structure authority, scope, data, audit and revocation for user-aligned agents. The point is simple: a reader-side filter that cannot be inspected or overruled merely relocates opacity.

The implementation challenge is substantial, but not exotic. Standards already exist in partial form. Browser vendors know how to display security signals. Enterprises already run policy engines over data flows. The novelty lies in bringing these pieces together around the individual as principal.

The Position Worth Holding

The temptation in every wave of digital disorder is to promise a purge: a detector that will finally separate authentic from synthetic, worthy from unworthy, human from machine. That promise is emotionally satisfying and operationally false.

Slop will not be filtered out of existence. It is the natural by-product of abundant generation meeting monetised attention. As long as publishing remains cheap and some channels remain open, the supply will continue.

The better question is not how to stop low-value content from being made, but how to stop it from paying. That requires two shifts. First, from guessing at origins to checking provenance, so that claims of authorship, review and modification become legible and attributable. Second, from platform-side ranking optimised for engagement to reader-side curation optimised for the user’s own standards.

Neither move is magical. Provenance does not create truth. Personal curation can still be abused, captured or narrowed into complacency. Adoption will be slow and uneven. Yet these approaches have something the rhetoric of detection does not: a mechanism.

They do not ask the internet to become cleaner by decree. They ask it to become less rewarding for sludge, and more navigable for people with standards. In an information economy shaped by abundance, that is the more serious ambition.

The death of slop, if it comes at all, will not look like victory. It will look like diminished returns, better filters, clearer accountability and a reader who is no longer the least powerful actor in the chain.