The Skimpflation of Everything

First the price went up. Then the packet got smaller. Now the packet is the same size and the price is the same, and what is inside is worse: fewer staff at the counter, thinner fabric, a support line that is a chatbot. Economists call this last stage skimpflation, and it is harder to see than the other two because nothing on the label changes. I want to argue that it has escaped the supermarket. It is happening to our buildings, our screens, our sentences, and artificial intelligence is the first machine built to do it at scale.

What is skimpflation?

Skimpflation is a price rise you pay in quality instead of money. The price stays where it was, the size stays where it was, and what you get for them gets worse. The word was put forward in October 2021 by Greg Rosalsky of NPR's Planet Money: "We propose a new word to describe this stealth-ninja kind of inflation: skimpflation." His definition is still the cleanest I know: "It's when, instead of simply raising prices, companies skimp on the goods and services they provide."

His examples were small and instantly familiar. "Airlines are putting customers who call them on hold for hours." At highway rest stops, "Napkin, utensil and condiment dispensers were empty." A hotel "was now skimping on its hot-breakfast buffet as well as maid service." Nobody was robbed. Everybody got a little less, and nobody was told.

Each of the three stages hides better than the one before.

Can official inflation statistics measure skimpflation?

For a shrinking packet, yes. For a thinning service, mostly not, and the statisticians are frank about why. In February 2023 the Bureau of Labor Statistics published an explainer by Kari McNair called "Getting less for the same price?" When a product shrinks and its price holds, she writes, "the CPI would include this price increase." The bureau even put a number on it: over 2015 to 2019, "product size changes (upsizing and downsizing) increased the CPI all commodity and services index by 0.01 percent per year." A smaller packet can be weighed, so it gets counted.

Quality is another matter. The bureau's page on quality adjustment explains that the index follows "a fixed basket of goods and services through time," and when an item vanishes, "a similar replacement item is selected." For some items hedonic methods remove "any price differential attributed to a change in quality." Then the page lists which method applies to which item, and the list is revealing. Cable television, wireless phone service and internet access get hedonic treatment. For "Full service meals and snacks", "Airline fares", "Hospital services" and "Physicians' services", the method listed is "none". The places where skimpflation lives, the waiter, the seat, the nurse's time, are largely the ones the list leaves blank.

I do not read this as negligence. Measuring whether a meal was served with care is genuinely hard, and Alan Cole, a former senior economist at the US Congress's Joint Economic Committee, said as much to Rosalsky: "It's just super-hard to systematically account for changes in quality when measuring changes in prices." Looking at official data that showed a bigger, more productive economy than in 2019, he went further. "That kind of suggests that the goods and services we're consuming now are better than they were before the pandemic," Cole said. "And I don't think that's true."

So skimpflation is not only hidden from you at the counter. It is partly hidden from the instruments a society uses to know how it is doing.

Why does skimpflation damage trust more than a price rise?

Because it breaks the meaning of the price itself. A price is a short sentence a seller says to a buyer: this much, for this. Inflation changes the first half of the sentence, openly. Shrinkflation changes the second half, but in print on the packet, where you can catch it. Skimpflation leaves the sentence word for word the same and quietly changes what the words refer to.

Once you have noticed this a few times, something shifts. You stop trusting the number. The hotel rate, the airline ticket, the subscription, the bill at a restaurant you used to love: each becomes a question rather than an answer. You read reviews looking for signs of thinning. You call the support line and wonder whether a person will ever pick up. You hold the shirt up to the light.

That vigilance is the real cost, and nobody counts it. The worse breakfast is a loss you can see. The time and suspicion every buyer now spends checking whether a price still means what it meant is spread across everyone, and it falls hardest on people with the least time to check. Prices that carry reliable information are a public good. Skimpflation spends that good down, a little at a time, and leaves no receipt.

There is already a word for this on the internet. In January 2023 Cory Doctorow described the life cycle of online platforms on his blog Pluralistic, in a passage the American Dialect Society later quoted: "Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die." He called the process enshittification. The American Dialect Society made it its 2023 Word of the Year. In November 2024 Australia's Macquarie Dictionary chose it too, with a definition that reads like skimpflation in a different coat: "the gradual deterioration of a service or product brought about by a reduction in the quality of service provided". Its committee said the word "captures what many of us feel is happening to the world and to so many aspects of our lives at the moment." People were not reaching for a joke. They were reaching for a name for the feeling that the price was the same and the thing was not.

Why does everything look the same now?

Because the pressure that thins a service also flattens a style: when the safest, most tested version wins, the average becomes the default, and the decision that made a thing particular is the first ingredient to go. The strategist Alex Murrell gathered the evidence in an essay called "The Age of Average", which circulated widely in 2023. His thesis is that "creative fields have become dominated and defined by convention and cliché", and his verdict is four words long: "Distinctiveness has died."

You will recognise every one of his examples. The café, in a style he notes is known as the Brooklyn Look or, in the journalist Kyle Chayka's term, AirSpace; of two such interiors he writes, "You might not even know you're in different spaces." The car: he quotes Jim Carroll, a former chairman of BBH London, remembering the day in 1983 "when we woke up and noticed that the cars all looked the same." The reason was simple. "They'd all been through the same wind tunnel." The building, through Rem Koolhaas's 1995 essay "The Generic City": "Convergence is possible only at the price of shedding identity." The logo: "Companies like AirBnB, Spotify and eBay have all dropped colourful logos with expressive typography". The film poster, where "every genre of film sticks to a relatively narrow set of clichés, codes and conventions".

Chayka carried the argument into his book Filterworld: How Algorithms Flattened Culture, published in 2024. Its publisher's description puts the café in one line: "The algorithm is present in the familiar neon signs and exposed brick of Internet cafes, be it in Nairobi or Portland." It also names the trade that worries me most: "What happens when shareability supersedes messiness, innovation, and creativity."

Sameness has more than one cause, and not all of them are bad. Cars met in a wind tunnel because air resistance is the same everywhere, and a more efficient car is a better car. Some convergence is just the discovery of what works.

But much of it is skimpflation by another route. The café still charges what a café charges. The new logo still cost what a rebrand costs. What has been removed is not material. It is the decision: the moment when a particular person chose something because it was right for this place rather than because it had already performed well somewhere else. The averaged café is not cheaper to sit in. It simply has less of anybody in it, and no label tells you.

What is intelligence inflation?

Intelligence inflation is what happens when competent text, images and answers become nearly free to produce in unlimited amounts: the quantity of "intelligence" in circulation explodes, and the value of each unit falls. In Turkish I call it zekâ enflasyonu. The analogy with money is close enough to be useful. When a central bank prints far more money than the economy can absorb, each note buys less. The notes are not counterfeit. There is simply too much of them for any one note to carry its old weight.

Generative AI does this to the things we used to treat as signs of a mind at work. A fluent paragraph, a polished image, a plausible answer: for most of history each cost someone time and attention, and so each carried a small guarantee that someone had meant it. That guarantee was the value. Now the paragraph is often fine, sometimes very good, and almost free. What it can no longer buy is what it used to buy: your attention, your trust, the sense that a person stood behind it.

The scale is no longer a guess. In May 2026, researchers at Graphite published a study of a random sample of 55.4k English-language article URLs from Common Crawl, published between January 2020 and March 2026, classified by averaging three AI detectors. Their finding: "In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026." They add that "since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%." The authors are candid about limits: human-edited, AI-assisted articles were not evaluated, newer models may be harder to detect, and the study does not measure whether these articles are read as often as human ones. Even so, by their count, roughly every second new article in their sample is now mostly machine-made.

The residue has a name. In May 2024 Simon Willison argued, crediting a post by @deepfates, that if content is "mindlessly generated and thrust upon someone who didn't ask for it, slop is the perfect term for it." In 2025 the Macquarie Dictionary, a year after crowning enshittification, chose "AI slop" as its Word of the Year, defined as "low-quality content created by generative AI, often containing errors, and not requested by the user." Two years running, an Australian dictionary picked a word for things getting worse. I find that more telling than any chart.

Why does AI produce the average?

Because it is built to. A generative model learns the patterns in a vast body of existing work and produces what is likely given what came before. The likely is, nearly by definition, close to the middle. The odd phrase, the strange colour, the idea nobody else would have had are the unlikely parts, and a system that rewards likelihood tends to round them off. I do not think better engineering will simply remove this. It is close to what the machine is for.

Two pieces of research show the shape of the problem. In a 2024 experiment published in Science Advances, Anil Doshi and Oliver Hauser gave some writers story ideas from a large language model. "We find that access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable, especially among less creative writers." That is real, and I return to it below. But the abstract continues: "However, generative AI-enabled stories are more similar to each other than stories by humans alone." The authors call it a social dilemma: "With generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced."

The second is about machines learning from machines. In Nature in July 2024, Ilia Shumailov and five co-authors reported that "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear." They called it model collapse. The tails are exactly the unusual parts: the rare word, the minority view, the distinct. Their result concerns a specific condition, training on generated data without care, and I am not claiming the web is collapsing. But when about half of new articles in a large sample are mostly machine-made, that condition is no longer hypothetical. An engine of the average, fed on its own output, gets more average.

Is AI the first machine built to do skimpflation at scale?

Yes, and I mean that precisely. Earlier machines made things cheaper and, sooner or later, the price showed it. Skimpflation's signature is that the price does not move. The packet looks the same, the label says the same thing, and only the inside changes.

Generative AI is the first technology that can remove the most expensive ingredient in almost any service, a person's attention, without changing anything the buyer can see. The support line keeps its number and its opening hours, and the voice on it is no longer anyone's. The article keeps its masthead and a byline-shaped space at the top. The invoice stays the same while the drafts arrive in seconds. The product description, the cover letter, the condolence note, the stock illustration: each can now be delivered at the old price, in the old wrapping, with far less human care inside.

I want to separate absence from accusation. Using AI is not skimping. Plenty of organisations use it to do more for the same money, or the same for less, and say so. Skimpflation begins at a specific point: when the price and the promise stay where they were, the hand leaves, and nobody tells the buyer. The undisclosed swap is the defect, not the tool. But the tool makes the swap easier than it has ever been, at a scale no understaffed hotel ever managed.

Wasn't everything always "better before"?

No, and an honest version of this argument has to start there. The complaint that standards are falling is among the oldest there is, and it is often wrong. Alan Cole, the economist who now worries about hidden quality loss, used to argue the opposite. He once thought, as he told Rosalsky, "that the world was getting better, faster than our official statistics would suggest," pointing to things like cheaper music streaming. For a long stretch he was right. A great many goods are better and cheaper than they were.

AI belongs on that side of the ledger too. In a study of 5,179 customer support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to an AI assistant raised productivity "by 14% on average", "including a 34% improvement for novice and low-skilled workers". Their evidence suggested the tool "disseminates the best practices of more able workers". Doshi and Hauser found the benefit especially among less creative writers. For a small shop owner who could never afford a designer, or a student writing in a second language, AI is not skimpflation. It is a step up, sometimes a big one. Raising the floor is a real good, and anyone who sneers at it has probably never had to stand on the floor.

So here is my answer. Both things are true, at different levels. AI raises the floor and narrows the commons. The same research that shows it helping beginners found "minimal impact on experienced and highly skilled workers", and the shared pool of stories getting more alike. The test I would apply is not whether a technology is new, or whether things were better in some remembered year. It is the trade. When AI gives you more for the same price, that is progress, and I welcome it. When it gives you less for the same price under the same label, that is skimpflation, whatever we call the tool.

Where can value still live?

In whatever cannot be printed. When a currency inflates, people move their savings into things that cannot be produced at will. Intelligence inflation will push value the same way: toward effort, toward a signature, toward the accountable human hand, and toward the distinct over the average. None of these can be generated for free, and that is exactly why they will matter more, not less.

Effort is scarce by construction: an hour someone spent on your problem is an hour they did not spend on anything else. A signature is scarce because it can be held to account, a name that will answer for an error and still be there next year. The distinct is scarce because the machine is pulled toward the middle and the tails are where it thins out. Shumailov and his colleagues reach a version of this from the other direction, writing that "the value of data collected about genuine human interactions with systems will be increasingly valuable." The record of real people doing real things is becoming the rare input. So is the real person.

That gives me a plain list of what I intend to keep paying for. A restaurant that still has enough people on the floor. A support line where a person can be reached, and says so. Writing with a name on it, from someone who will answer if it is wrong. A designer who will tell you the template is the wrong answer for your street. A building that could only stand in the town it is in.

And I would ask sellers for one disclosure, the way packets list their ingredients: who made this, who checked it, and who answers if it is wrong. A vendor could adopt that line tomorrow, and a buyer could look for it. It would ban nothing. It would only stop the swap from being silent, which is the part of skimpflation that does the damage.

Murrell ends his essay with advice that works for everyone: "When the world zigs. Zag." The average is about to become free, endless and everywhere. That makes it the cheapest thing in the world, and the particular, the meant, the signed and the cared for, the most valuable. Skimpflation takes the care out and keeps the price. Our answer is to notice, to ask, and to keep paying for the hand.

---

This is an opinion piece. It argues a position and it is signed. The record comes from: Greg Rosalsky's Planet Money article proposing skimpflation (read through NPR's text edition), including Alan Cole's remarks; the US Bureau of Labor Statistics' quality adjustment page and Kari McNair's Beyond the Numbers article on shrinkflation; Cory Doctorow's Pluralistic post and the American Dialect Society's 2023 Word of the Year announcement; the Macquarie Dictionary's Words of the Year for 2024 and 2025; Alex Murrell's "The Age of Average" and Kottke's April 2023 link to it; the publisher's page for Kyle Chayka's Filterworld; Graphite's May 2026 study of AI-generated articles; Simon Willison's post on slop; Doshi and Hauser's study in Science Advances, read through its Europe PMC abstract record; Shumailov and colleagues' paper in Nature; and Brynjolfsson, Li and Raymond's Generative AI at Work. Where the essay leaves the record and enters argument, it says so. Our writing workflow runs on Anthropic models. The interpretation is mine.