A hypothesis on the not-so-distant future of software, books, music, and movies, in which most of what we consume gets cheaper, more abundant, and noticeably worse, while the human-made variant moves into a luxury segment of its own.

Disclaimer: This is an opinion piece and most of it is speculation about a future that has not arrived (yet?), based on a few data points that have. As usual, summary at the end.

A few years ago I would have laughed at anyone telling me that there is a serious market for ten-dollar drills, two-dollar dresses, and one-dollar pairs of shoes shipped from a warehouse on the other side of the planet. Today, however, that market exists and it has a name, and it is even publicly traded (sort of, through holdings). TEMU, Shein and a few others have built frankly mind-boggling businesses around the idea that if you make production cheap enough, fast enough, and just barely good enough to look right on a phone screen, an enormous part of the population will buy it, even when the product breaks within a week, when the materials it is made of contain worrying levels of toxic substances , and when the carbon footprint of one delivery exceeds that of an equivalent local purchase by orders of magnitude.

The key to this sort of business model is not innovation, but instead the externalization and compression of cost. Somewhere upstream, people work seventy-five hours a week , in conditions most readers of this website would refuse to even visit, so that the rest of us can have a cheap plastic spatula at our doorstep within five business days. While the visible price collapses, the invisible costs get distributed onto landfills, lungs, and ultimately people that we will never meet.

What follows is a hypothesis I cannot prove but have been turning over in my head for a while, as we are watching the same thing happen to software, books, music, (film-)scripts, and most of the digital goods and services we consume. The cheap labor in this case is not human, it is a Large Language Model (LLM), or what many people these days call “AI”, and the externalized cost is, among other things, quality, which requires craftsmanship to produce, and attention to perceive. And just like with physical goods, we will probably end up with a two-tier market, in which we have a large and massively profitable lower tier of generated slop, and a smaller, more expensive upper tier of work that is still recognizably human.

I’d like to call this the TEMU-fication of software, digital goods and services, and describe what it might look like.

Cheap labor

For decades, the global fashion industry has relied on a workforce that has almost no leverage and no voice, and for which the economics work because someone, somewhere far away, will sew a t-shirt for less than the price of a coffee. Without that skewed arrangement, the entire fast fashion business model collapses. The garment in your hand is only cheap to you because it has been expensive to someone else, in ways that the price tag does not show.

Modern Large Language Models occupy a similar position in the economy, with one important difference, which is that there is no human being in the sweatshop, only a stack of GPUs trained on a corpus of work that other human beings produced over the course of decades. The labor that has been compressed is historical and the model is a kind of compressed copy of the work of millions of programmers, writers, illustrators, and musicians, served back at near-zero marginal cost. Well, at least in theory, and only if the hyperscalers find a way to lower the cost per token, but that’s a different topic.

However, the result is the same. A class of goods can suddenly be produced for an order of magnitude less than before. And, just like with TEMU , those goods turn out to be just barely good enough.

Vibe-coded software

The most direct manifestation of this so far is what is being called vibe coding. The term refers to the practice of describing what you want in natural language to an LLM, accepting whatever it produces, iterating over it with more refined descriptions of the basic idea and eventually shipping the result into production. Whether the developer actually understands what was generated is increasingly considered an implementation detail. And while the output is technically software, the question is what kind of software it is.

A 2025 Veracode report found that approximately 45% of AI-generated code samples failed security tests and contained critical vulnerabilities from the OWASP Top 10, and a multi-language, multi-model academic study that evaluated outputs from Claude, Gemini, Codestral, GPT-4o and Llama-3 across Python, Java, C++ and C, found that a substantial fraction of generated snippets were either non-compliant with basic secure coding standards or actively triggered classified weaknesses (buffer overflows, hard-coded credentials, SQL injection, cryptographic misuse, path traversal, you name it). Even more concerning is a peer-reviewed 2025 paper from IEEE-ISTAS that documents a 37.6% increase in critical vulnerabilities after just five iterative prompts, suggesting that the more you let the model refine its own code, the worse the security posture gets.

When these issues compound over time, the result is a higher total cost than traditional development. However, this doesn’t matter when you don’t think long term, but fast fashion instead. Also, none of this is to say that an experienced engineer cannot use these tools well, because they certainly can. The issue is what happens when the same tools are used by someone who does not know what good looks like in the first place, and there is nobody downstream of them who does either. The output passes the basic test of it runs and looks plausible, ships into production, and accumulates the kind of architectural and security debt that surfaces only when something goes very wrong .

Note: There are credible voices in the industry, particularly from the AI tooling vendors themselves, who argue that AI-assisted development raises a floor more than it lowers a ceiling. In this view, the median piece of software has always been mediocre, written under deadline pressure by tired humans, copied from Stack Overflow without much thought, and held together by duct tape. If an LLM produces output of roughly comparable quality in a fraction of the time, the argument goes, nothing got worse. We are simply removing a bottleneck.

I find this argument partially persuasive, and partially convenient for the people making it. It is true that a lot of software was already not great, but it is equally true that there is a difference between bad code written by a human who at least understood what they were doing, and bad code written by a system that does not understand anything. The first kind can be questioned and corrected, but the second kind tends to compound, because the person shipping it cannot answer why it does what it does. At least for now.

Vibe-written books

Software is not the only place where this is playing out. The book industry is arguably further along, with estimates suggesting that somewhere between ten thousand and forty thousand AI-generated books are uploaded to Amazon’s Kindle Direct Publishing platform every month, many without any disclosure that a model was involved. In June 2023, the Kindle Top 100 bestseller list was found to contain only 19 books written by humans . Amazon has since introduced limits and disclosure requirements , but enforcement is patchy and authors continue to push back against what looks like a slow flood.

Categories that have been hit particularly hard include travel guides (generated guides to cities the author has never visited, with restaurant recommendations that don’t exist), nutrition and health (generated diet advice with citations to studies that don’t exist), and public-domain rewrites (generated adaptations of older books, relying on the recognizability of titles that the actual authors never agreed to). Travel guides in particular have produced a small genre of stories where readers arrive at addresses that turn out to be empty lots, or follow walking directions through neighborhoods that no human would ever recommend.

Note: The defense, again, is that the bottom of the book market was always full of filler, that print-on-demand has been around for a long time, and ghost-written business books and assembly-line genre fiction predate generative AI by decades. However, the new thing is the scale at which low-effort content can now be produced, and the speed at which it can drown out the rest of the catalogue. Authors are competing for shelf space against entities that can ship a hundred new titles in a weekend.

Vibe-created articles & video

A 2025 analysis of 65,000 English-language articles published since January 2020 found that a little over half of all new articles on the internet are now AI-generated , and it’s not only the written word that’s being churned out by machines. YouTube has its own version of the problem, where, according to a Guardian analysis, nearly 10% of the world’s fastest-growing channels feature nothing but AI-generated content , and on Shorts specifically more than one in five videos served to a new user is low-quality AI-generated material .

Vibe-produced music

Not even the highly creative and (up until recently) human process of making music is immune to this TEMU-fication. Spotify has been removing ghost artist tracks for years, but the practice scaled up dramatically when generative tools made it trivial to produce convincing lo-fi background music in arbitrary volume. The platform has reportedly removed 75 million spammy tracks in a single year , and high-profile acts like the AI-generated band The Velvet Sundown amassed over a million streams before being unmasked. There has been at least one criminal case, involving over $8 million in fraudulent royalties , built entirely on AI-generated music and bot streams.

However, that is no reason to applaud Spotify , as the company appears to fight the AI spam only when it’s someone else trying to make money off of it.

Pushback… or delay?

However, there is a sliver of hope, as engagement with AI-generated articles reportedly dropped by around 40% in 2024, and human-generated content seemingly still gets roughly 5.4× more traffic than AI-generated material in some studies. About 38% of consumers openly express skepticism about AI-created content, and people do still seem to be voting with their attention. Whether that vote is powerful enough to shift incentives at the platform level is a different question, and personally I’m not particularly optimistic, especially given that the platforms profit either way.

Let’s take Netflix as an example. From my understanding, the WGA’s 2023 deal explicitly prevents studios from treating AI-generated material as source material, or from using AI to write or rewrite scripts, and Netflix was seemingly bound by that agreement until at least May 2026. Netflix’s own Generative AI Production Guidelines also seem to reflect this, stating that AI is permitted in ideation, but that its use should not replace or materially impact work that would otherwise be done by union-represented writers, actors, or crew members, without proper approvals. While that sounds reassuring on the surface, it is, in my view, a delay and not a limit.

The same company has publicly committed to going all-in on AI in its production pipeline , has signed deals with VFX automation providers that explicitly put a chunk of the global VFX workforce at risk, and has already used generative AI in at least one of its programs ( El Eternauta ). The trajectory seems to be “use AI everywhere it is contractually allowed right now, expand into the rest the second the contracts permit it, and spin the result as dEmOcRaTiZaTiOn Of CrEaTiViTy”.

Looking forward

So here is my specific (and quite possibly wrong) prediction:

Within the next five to ten years, Netflix will offer a basic subscription tier whose catalogue consists predominantly of AI-generated or AI-assisted content. We are talking generated procedural shows where each episode is remixed from a small set of templates, generated kids’ content that is vaguely educational and impossible to remember an hour after watching, and generated dramas that recycle plots from existing IP and vibe the rest. For this, the viewer pays the lowest monthly price, while the platform pays nearly nothing in production cost and keeps an enormous margin. The only “upside” for consumers will be the lack of ad breaks, as targeted advertising will quite possibly be injected in real-time into the show you’re watching, seamlessly blending into the storyline without you noticing it, but ultimately still triggering your ape brain to crave a refreshing soda or a sweet treat .

Their premium tier, meanwhile, will become the human-made tier. Series with credited human writers, films with credited human directors, and performances by humans whose likeness has not been digitally replicated. The marketing will not call it human-made, because that would be admitting that the cheap tier isn’t, but the price difference will make it obvious. You will pay extra for the same thing Netflix has been selling you all along, except now it is positioned as a luxury.

Clearly, I cannot prove that this is what will happen. Netflix’s own guidelines, as written, prohibit it, and the WGA deal forced a delay. But once the contractual block has lifted, the financial logic is hard to argue with. A streaming service that can produce good enough content for fractional cost will eventually try to. And, mind you, Netflix is just one example. The same logic applies to every other content-distribution business with a subscription model and a margin.

Handmade goods

If you want to know what the human side of this two-tier world looks like, I think the best existing model is the handicrafts and handmade goods market . By 2025, that market was estimated at roughly USD 987 billion globally, with projections reaching over USD 1 trillion by 2035 . There is data suggesting that U.S. consumers already spend almost a fifth of their money on handmade goods rather than on mass-produced equivalents, and over half of handicraft buyers globally indicate a preference for products that are eco-certified or made from natural materials, going in the exact opposite direction of what TEMU has been doing.

What this market shows is that industrialization does not erase the artisans, but pushes them into a different segment. People did not stop buying handmade chairs when factories started making chairs cheaply. While the masses opted for the cheaper, mass-produced items, a small but sustained minority of buyers continued to seek out the human-made version, and over time were willing to pay a premium for it.

If the hypothesis holds, software engineering, writing, acting, illustration, composition and the other content-producing professions will undergo something similar. The bulk of the market will migrate to the cheap, mass-produced, generated tier, while a smaller market will continue to value, and to pay for, work that is verifiably the product of a thinking, breathing, opinionated human being. We are already seeing the first signs of this in agencies that explicitly advertise human-only content (at a premium), and in licensing companies flagging tracks as human-composed to distinguish them from AI library music.

I think that the interesting question is not whether this segmentation will happen, but what proportion of the market ends up in each tier, and how robust the upper tier turns out to be.

Ultra-processed

There is a darker version of this analogy. Roughly 57-60% of the daily caloric intake of the average adult in the United States and the United Kingdom now comes from ultra-processed foods . Across 22 European countries the share ranges from 14% to 44% , depending mostly on how protected the local food culture has remained. These foods are cheap, abundant, available everywhere, and nutritionally inferior to the alternatives in ways that have been studied at length . People know this, but they eat them anyway, often because the alternatives are slower, more expensive, harder to find, or require skills that have not been taught.

I suspect that AI-generated content is on the same path. The cheap tier will not be a marginal phenomenon serving a marginal audience, but it will be the default, the cornerstone of how most people consume software, entertainment, news, and information, because it is what the platforms will serve them and what their monthly subscription covers. Some will care enough to seek out the alternative, but most will not, just as most people, knowing what they know about ultra-processed food, do not change their grocery habits.

But maybe I’m wrong

Probably the strongest counter-argument to all of this is that LLMs are still early, that the quality issues are transient, and that within a few model generations the gap between AI-generated and human-generated work will narrow to the point where the distinction stops mattering or might not even be possible anymore. If that is true, the two-tier picture collapses, because there is no longer a quality difference to justify the upper tier, only a marketing difference. The handmade analogy breaks because, unlike a hand-built chair, a generated novel is functionally identical to a written novel once you can no longer tell them apart.

However, I am doubtful that this is going to be the case. There are tasks where I have watched the gap narrow faster than I expected, but there are also tasks where the gap has stayed stubbornly fixed and the failures have just gotten more sophisticated. My instinct is that for narrow, well-bounded technical work, the gap will close further. For long-form work that depends on a coherent worldview, lived experience, and, most importantly, emotions, I doubt it will, because the model has none of those.

The second counter-argument is that the consumer backlash will be stronger than I am giving it credit for. The 40% drop in engagement with AI-generated articles is not nothing, and platform incentives may shift if users start to penalize AI-flooded feeds. Apple and others have started experimenting with content provenance and disclosure schemes that, if widely adopted, could stop the worst of the flooding. So it is possible that I am underestimating the immune response.

The third counter-argument is, that the cheap tier might not be sustainable at all, because AI-generated content trained on AI-generated content degrades model quality , and the broader ecosystem ends up poisoning its own training data. If that turns out to be the dominant dynamic, the cheap tier could collapse before it becomes entrenched.

I think all three of these arguments are valid and have a certain weight to them, but none of them are strong enough, in my view, to make me confident that the TEMU-fication will not happen. They might modulate how it happens, but they probably do not stop it.

Summary

Initially, I went looking for an optimistic ending for this write-up, to say that software engineering is not going away, and writers are not going away, and actors are not going away. And while all of that is, I think, true, none of it should be confused with things will look the same.

What I expect, and what I am to some degree already seeing, is that the people producing software, books, music, scripts, and other human-made work will not disappear, but they will get pushed into a narrower, more specialized, more “luxury”-coded part of the market, pretty much the same way hand-bound notebooks, independent record stores, and small bakeries that mill their own flour did. There will still be a livelihood in it, at times a very good one, but it will look vastly different, and there will probably be fewer people making a living in these fields. My assumption is that they will be more visible inside their niche, but less visible outside it, and they will make their case in part on the basis of provenance, where something was made by a human who knew what they were doing, and you can tell.

Meanwhile, the bulk of what most people interact with will, I suspect, be generated. Some of it will be fine, and some of it will be ultra-processed, in the same sense that a frozen lasagna is ultra-processed. It will be functional, calorically adequate food, but it will not be what your Italian grandmother was making. People will nevertheless eat it because it is there, it is cheap, it is convenient, and because the alternatives have been priced out of their daily life.

There is no “inevitability” to it, because none of this is really decided yet. There are still choices, made by platforms, by regulators, by consumers, and by the people doing the actual work, that will shape which tier ends up being how big and how durable. The handmade market exists because enough people kept buying handmade goods to make it viable. The human-made tier of software and digital goods will exist because enough people keep buying it, or it won’t exist at all. If you are someone who writes code, or stories, or music, or scripts, by hand, with intent, and with a point of view, I do not think the LLM is going to kill your job. I do think, however, that it is going to change the shape of the market you operate in, push you toward the upper tier (whether you wanted to be there or not) and ask you to make a more deliberate case for why your work is worth the difference in price.

For the rest of us, the more interesting question is which tier we are choosing to consume from, and whether we are choosing it on purpose, or just because it was what the algorithm served us by default. I have my suspicions about the answer, but I would love to be wrong.

Enjoyed this? Please consider supporting my work.

Published on 2026-08-02 and updated on 2026-08-02 in journal and tagged with ai art business society