This page is analysis, not a forecast, and the distinction matters on a subject where confident predictions are cheap. What follows is three paths that are already partly visible, the signals that would distinguish them, and an honest account of what is not known.
Path one: the flood continues and reading gets colder
The cheapest outcome, and the one currently most visible.
Slop production stays cheap, platform incentives do not change, and readers adapt by trusting almost nothing by default. The adaptation is already happening. People skim for a specific fact and leave. They retreat to a small set of sources that have not burned them yet. They stop using channels that have become unusable, which is what happens when a marketplace category fills with listings that do not correspond to real products.
The cost of this path is not that the internet stops working. It is that the default assumption shifts from provisional trust to provisional distrust, and the readers who pay most are the ones with least time to verify.
Research on the term's normative status locates the real harm in the aggregate: no single item has to be bad for the total to raise the cost of knowing anything. The concern is measurable in the one place it has been measured, which is the workplace, where the cost of checking unchecked AI output runs to nearly two hours per incident.
Path two: provenance and ranking make slop expensive
The intervention path, and the one with the clearest mechanism.
Two levers would change the economics. The first is provenance: verifiable signals attached to media recording how it was made, so that synthetic material can be identified without anyone having to guess. The second is ranking: platforms deciding that unreviewed bulk content does not deserve distribution.
A Columbia University report on AI slop names both, framing the debate as being about redesigning ranking algorithms and provenance infrastructure rather than about content bans. That framing is worth taking seriously because it identifies the actual constraint: bans are unenforceable at this volume, while ranking and origin metadata are decisions a small number of organizations can make.
The obstacle is incentives rather than feasibility. Provenance standards only work if enough producers adopt them, and the producers with the strongest reasons not to adopt are the ones creating the problem. Ranking changes cost engagement in the short term, which is exactly the metric platforms are judged on.
There is a small piece of evidence that institutions will move when pushed. After complaints about unchecked model output in research, arXiv implemented a rule restricting submissions that show clear evidence the authors did not check their generated results. Commentary on that decision noted the rule is narrower than the headlines suggested: it addresses the quality of the submission rather than the use of a tool.
That narrowness is instructive. The enforceable version of the rule is almost always about review and accountability rather than about tools, because that is the part that can be checked.
Path three: verification becomes a product
The market path, and the most speculative of the three.
If checking costs rise, someone will sell the checking, or sell content that has already been checked. Subscription newsletters, membership-supported publications and curated recommendation services are all versions of this, and several already market themselves on the fact that a person wrote and verified the thing.
If this path develops, the likely shape is a two-tier internet: free material where provenance is unknown, and paid material where someone vouches for it. The uncomfortable implication is that being able to trust what you read becomes something you buy, which is a worse outcome than nearly every alternative and also an entirely plausible one.
What is genuinely unknown is whether readers will pay. The demand for verified human work has never been tested at this scale, because the alternative was never this cheap. Anyone who tells you they know the answer is guessing.
The signals that would tell you which path is happening
Four things are checkable, and they are better than opinion:
- Whether provenance metadata becomes common in ordinary media. Quiet adoption rather than announcements. If generated images routinely carry verifiable origin data within a few years, path two is underway.
- What ranking systems reward. If a major platform publicly changes what it distributes and holds the line through a bad quarter, that is the strongest available signal.
- What models are trained on. If major labs start paying for licensed human work because the open web is unusable, the contamination argument has been confirmed by the people best placed to know.
- What readers do. Subscription growth for human-verified publications, and abandonment of search for specific categories, are both visible from outside.
What is not known
Three things, stated plainly, because the alternative is pretending to a certainty nobody has.
Whether audiences care. Consumption of slop is enormous. People watch the generated videos, enjoy the absurd images and share them. Any account that assumes a general disgust is assuming its own conclusion.
Whether detection improves. The most careful attempt to build a measurement framework for slop in text concluded that fully automated scalable detection remains an open problem, and that expert annotators disagree on borderline cases. If that changes, several of these paths change with it.
Whether the aggregate harm is as large as it feels. The mechanisms are documented. The magnitudes are not. A site that writes about slop should be the last one to substitute a feeling for a number, which is why everything above is labeled as a path rather than a prediction.