Applied AI Academy

Cheap generation does not remove cost — it moves it to attention, trust and skill

AI slop, enshittification, model collapse and cognitive offloading: the harms that arrive with no villain and nobody to fine

The question: Which AI harms arrive without anybody deciding to cause them — and what can a person or a firm actually do about those?

This module has covered the harms that have an owner: a biased model, an unlawful deployment, a breach. Each comes with a control and usually a regulator. This lesson covers the category with no owner at all — what ordinary incentives produce once generating costs almost nothing. It is already visible in your feed, in the products you pay for, in the training data of the next model, and in the uncomfortable one: what you are still able to do without help.

What the lesson covers

Start with the economics, because the moral framing gets in the way. Herbert Simon put it in 1971: information consumes attention, so a wealth of information creates a poverty of attention. Generative AI drops the cost of producing information to near zero while leaving the amount of attention exactly where it was. Nobody has to intend any harm for that to end badly — every individual actor is behaving rationally, and the outcome is a commons that fills with plausible material faster than anyone can sort it. That is the shape of every harm in this lesson: not a decision, an equilibrium.

**AI slop** is the first of them, and it now has a name in the dictionary — word of the year for 2025 at both Merriam-Webster and the American Dialect Society. It is content generated in volume to capture attention rather than to say anything: the "Shrimp Jesus" images that farmed engagement across Facebook in 2024, the review sites that answer no question, the LinkedIn post that could have been written about any company. The business consequence is not aesthetic. When competent-looking output is free, competence stops being a differentiator — and what becomes scarce, and therefore valuable, is evidence, specificity, and a named human who can be held to what was written.

**Enshittification** is the same pressure applied to platforms. Cory Doctorow's 2022 formulation describes three phases: a platform is good to its users, then degrades their experience to favour its business customers, then claws value back from those customers too, and then it dies. It was the American Dialect Society's word of the year for 2023 and Macquarie's for 2024, which tells you something about how widely the pattern is recognised. AI accelerates it for a specific reason: recommendation, pricing and ranking become continuously tunable, so value can be reallocated in increments too small for any user to notice within a single session. If your product's quality is a dial wired to a quarterly number, somebody will eventually turn it.

**Model collapse** is the loop closing. Shumailov and colleagues showed in Nature in 2024 that training a generative model indiscriminately on generated content produces irreversible defects: the tails of the original distribution vanish first, so the model forgets the rare and the unusual before it forgets anything obvious. Two consequences for a business follow. Data of verified human origin appreciates, because it is the input the whole industry now needs and cannot manufacture. And a web corpus can no longer be assumed to be human, which makes provenance a data-quality dimension rather than a compliance nicety — a direct extension of Lesson 7.

The last one lands inside the reader. **Cognitive offloading** — handing a thinking step to a tool — is often free and sometimes obviously correct; nobody laments the mental arithmetic lost to the calculator. The evidence that it is not always free is real but should be read carefully. A Microsoft Research and Carnegie Mellon survey of knowledge workers found that higher confidence in the AI was associated with less critical-thinking effort, and higher confidence in one's own expertise with more. An MIT Media Lab study measured EEG activity across essay-writing sessions and found the weakest brain connectivity, the lowest recall and the lowest sense of ownership in the group that used a language model — 54 participants on one task, which is suggestive rather than conclusive, and it should be quoted that way. The mechanism underneath is older than either: automation complacency, documented in aviation for decades. The rule that follows is practical, not moralistic. Offload the step where verification is cheap. Keep the step you are paid to be good at.

Key points

Framework — Who pays: the four bills of cheap generation

Near-zero marginal cost does not remove cost, it relocates it. ATTENTION — the commons fills with plausible filler and sorting becomes the expensive part. TRUST — provenance turns into the scarce signal, so it has to be recorded at the moment of creation. PLATFORM — quality becomes a dial connected to a number, and dials get turned. SKILL — the step you stop performing is the step you stop being able to perform. Before any decision to generate at volume, name which bill it is charging and who receives it.

The lab

Measure the four bills where you can actually see them: your feed, a platform you depend on, your own data, and your own week.

Deliverable: Your one-page provenance and offloading policy: what you record when you generate, what you disclose, which steps you will not delegate, and the measure you will watch to know whether that is holding.

Open this lesson, its lab and its quiz

Sources and further reading