#4753: The Age of Ensloppification
There’s been a mixture of hype and controversy about this new AI agent thing called Openclaw1. It bills itself as a Personal AI Assistant and, basically, you set it up with a bunch of integrations and let it loose on your life. To me it appears to be like Home Assistant but with AI agents.
The usual cohort of early adopters and AI accelerationists have bravely submitted themselves as guinea pigs in this bizarre experiment and bought Mac Minis en-mass in order to wire Openclaw into their entire life.
It’s fascinating, like a car crash in slow motion. Like a Waymo-self-driving-taxi crash in slow motion. They are doing this so we don’t have to, but their experiment requires that they involve other people without their consent and that creates ethical concerns and externalities.
Of course, Openclaw isn’t the beginning of this and neither is it the end, so let’s see if I can break this down somehow and by the end we’ll get to the juicy stuff.
Dead Internet Theory
The internet might not be dead yet but it is certainly on life support as AI does what a Russian botnet can never do: democratise slop. Why even bother paying humans pennies on the dollar to operate a bot farm when you can get an AI subscription and a bunch of agents instead?
The signs of AI-infused slop are everywhere and even as the technology evolves you’ll start to see the patterns. On reddit in particular, the current trend is to post image macros that consist of a stock or AI generated image and a caption with keywords highlighted for emphasis (akin to how a newspaper would use capitalisation or scare quotes to do the same). The story invokes a feeling of incredulity and there is no source for the claim, so its only purpose really is to drive engagement on something that is most likely fabricated.
I ditched other forms of social media a long time ago and you could not pay me enough money to sign up for Facebook, Instagram, Twitter, or TikTok, so I can’t really speak to those unless they end up being shared on Reddit or LinkedIn. However, if the targeted misinformation posted on Facebook required substantial funding during the Brexit campaign back in 2016, it is now a trivial endeavour. So much for not believing everything you read.
This, I think, is the dawn of what I call the era of ensloppification, which I consider to be AI used exclusively in service of enshittification. It can only really get worse on the major platforms and walled gardens as they all have a vested interest in making AI happen, and as models recursively train on their own slop we’re going to see a continuing decline in creative output.
Optimistically speaking I hope this means that we put a higher value on artisanal productions, because those kinds of arts and crafts will become scarce as AI generated slop suffocates the market. I am more than willing to put my money where my mouth is on this topic.
Post-creative content
There was a time in the early twenty-tens where the word ‘content’ became in vogue as a catch all term for, well, arts and crafts. Music, videos, blog posts, podcasts, you name it…it was all gathered under the umbrella of content. And then the content creators came. I always found that to be a soulless definition of creative work that opened the gates to mindless consumption, because content is consumed while music is listened to, a film is watched, a book is read, and a painting is pondered over. In the later years we would describe a subset of this content as brainrot and the act of consumption as doomscrolling.
That started over ten years ago but AI turns the dial up to eleven. The threat is that it deflates the value of creative work to being beyond worthless as effort and elbow grease is replaced with vague promptery. Slopping a few vague talking points together in Suno and getting stereotypical country music out of it is not an artistic endeavour in any sense of the imagination, and it’s not made any more significant a contribution because you can go on TikTok and lipsync to it. Literally anybody can do that and get a result that is as mid as the average of all its parts.
“You can just not listen to that shit though, right?” you might ask, and maybe you’re right, but streaming services with a strong anti-AI policy are thin on the ground. Spotify doesn’t care about AI because it gets ad revenue; Youtube Music doesn’t care about AI because Google has ads, and it also has Gemini and it can’t be anti-itself. Maybe you’ve got Apple Music for now or you just go back to buying CDs and vinyl records. Unless a billionaire like Taylor Swift comes in and consumes the entire supply of physical media to release 50 versions of a song in a shocking-but-predictable act of commercial cynicism, showing that Taylor Swift isn’t an artist but a multinational corporation.
The second order effect with all of this is that it affects the livelihood of freelancers who lose out on work because someone can defer to an AI instead and get objectively worse but passable results. I’ve seen plenty of AI slop posters around in pubs and at market stalls that are absolutely nonsensical, and plenty of markets will try to hawk AI generated postcards now while charging as if they were hand-made.
Voice actors in games can be hit particularly hard as unethical clients record the actor’s voice in order to train an LLM to replace them. Or they might skip the voice acting entirely and get an AI service to do it instead. While this may serve a legitimate purpose in a prototyping stage or for personal hobby projects, I’m of the position that it’s not ethical to release this into production if you otherwise intended to hire people to do the job.
Sam Vimes’ Theory of Socioeconomic Unfairness
Physical goods are not exempt from this level of ensloppification either. The natural evolution of marketplaces like Etsy, Amazon, probably Vinted at some point, means that you will have to sift through pages and pages of slop to find legit products, either artisanal or sold by reputable brands and not random anagrams.
The Sam Vimes’ theory comes from Terry Pratchett and, rather astutely, posits that it is expensive to be poor. He wrote this before payday loans, Klarna, and PayPal Pay in 3 existed, but we see this play out as those with lower income are subject to lower recurring costs that in their totality amount to more than what it would have cost to get something more durable in a single payment. People in poverty needing to manage cash-flow.
All of this serves to erode your trust in buying online and, from a socioeconomic standpoint, I think those with the least money to spare will be hit hardest by this as they are misled by deceptively cheap offers. This creates an influx of low-level scams that make money through scale.
Post-truth journalism
1984 and Brave New World dream up a fearful future where the population is controlled either through fear and surveillance (Orwell) or through entertainment and hedonism (Huxley). Two sides of the same coin. The concept of HyperNormalisation2 is practically a prototype of what looks to be contemporary techno-feudalism, as it makes the case for constructing a parallel reality that doesn’t align with the real world as we experience it.
As you may be aware, Vladimir Putin’s Russia doesn’t play one side, or even both sides, it plays all of them to maximise division and chaos. This is an effective application of HyperNormalism as there is no single truth in play at one time, there are instead several of them. It muddies the waters and serves to exhaust you so you fall back to what is most plausible to your political stance or, ultimately, give up and become apathetic.
While Russia have become experts in this, we now have ensloppified journalism that serves the same purpose. The goal in the case of journalism is not necessarily to create division but to generate ad revenue, as the more outrageous their content is, the more money they get from page views and clicks. This is arguably one of the contributing reasons to Donald Trump being taken seriously, as treating him as the threat he is would generate less ad revenue from having less to report on.
AI makes this easy because journalists don’t even have to research any more, they can just scrape Reddit for vox pops. That are also mostly bot or AI generated.
There is no appetite for any media or media-adjacent business to tackle this because it brings them money. Meta would lose a huge amount of revenue if they attempted to stop this.
Non-consensual sexual material and CSAM
There is not a single word that can be said to legitimise or justify this, and I’ve gone back and forth about whether I even mention it in the context of slop, as if I’m trying to judge the quality of it. No, nope, fuck no. Given Twitter’s involvement in a lot of that, I sincerely hope the leadership is held to the highest accountability possible. Looking at you, Musk.
There is no legitimate justification for using AI to work around consent and do the unspeakable.
Vibe coding
My bias as a software engineer will become clear here, but that’s not to say I don’t have complicated feelings on the topic. While I’ve personally benefitted from using AI agents in my work to boost my productivity, I’m aware of the various trade-offs involved and I discussed some in a previous article3.
There is something to be said about making the field of software engineering more accessible, and AI does help with that, but it does so at the cost of education and the development of wisdom and experience. At best you will learn how to become effective with AI but your growth as a software developer will be limited if you stay hands-off.
I do wonder if the profession of software engineer will suffer because AI replaces the necessity of passing down knowledge, but I keep that opinion in check by considering that the profession has evolved drastically over the course of 70 years or so. Nobody writes programs with punchcards any more; knowledge of assembly language and CPU architecture is relevant but only in specific fields; embedded engineering, compiler development, high frequency trading/quants will demand proficiency in C or C++ or maybe even haskell or another ML; game developers will go for C# and web developers Typescript, Ruby, PHP, Python; data engineers will stick to Python. It’s ultimately a diverse field that continues to adapt because technology moves fast. This is nothing short of incredible and we as software engineers live and thrive in specialised abstractions.
However, vibe coding without any programming nous creates a lot of latent problems, or vulnerabilities. This isn’t much different from, say, me learning PHP back in 2004 and implementing SQL injection vulnerabilities and XSS all over my app because I didn’t know any better, and neither did the sources I was learning from.
The ensloppification in this context is essentially the demotion of solid engineering principles in favour of velocity, growing ARR, moving fast and breaking things, and other such startup jargon. The business is served by this approach in the short term but as they say, a Lannister always pays his technical debt.
It might take some time to see this play out in full but I imagine a number of startups starting strong, finding product market fit, and then realising that their entire budget is swallowed by the can they kicked down the road for months or years and their velocity slows to a crawl as firefighting takes over. Yes, I’ve been there.
Accelerationism
If you haven’t heard of this before, it’s essentially a belief that the only way to escape the current capitalist system is to run headlong into collapse. As with anything, there is a spectrum of interpretation based on where one stands politically or even technically or religiously, so you have various flavours of it based on where you sit either on the left, the right, or farther out on the fringes. I’m not going to go into it beyond this paragraph but you might connect this to, say, the current Trump administration, Project 2025, and otherwise the almost unanimous support for shovelling as much money as possible into AI. Of course, this desired collapse isn’t to our benefit, it’s to theirs.
Accelerationism and slop go hand in hand because the slop is a forcing function for eroding trust in the system. This sounds like a conspiracy theory and you’re probably thinking I should put on the tin-foil hat, but you have to consider the externalities of the typical AI startup now: the people hit hardest by its introduction are those that have to handle the [unsupervised] output of it, or see a machine doing the job they used to do but worse. The money sloshing around in the investment circles is completely decoupled from the reality on the ground as billions and billions trade hands behind the scenes.
I’m afraid I’m gonna lose people by going too much into this, but it sets the stage for Openclaw, which is the thing that triggered this entire post.
Autonomous agents
A few days a pull request was submitted to `matplotlib’5, which looked fairly innocuous at first glance. One of the maintainers rejected the change because they wanted to leave the issue for new [human] contributors to get a foot in the door and they figured out the entire thing was generated by AI by reviewing the contributor’s profile.
That sounds fair enough, doesn’t it? It was identified that the contributor was running Openclaw, and that software was acting on its own initiative to submit contributions and respond to feedback. In this situation, it responded badly6 with a targeted hitpiece before following up with a trite apology that still implied a conflict7 and then, somehow still holding a grudge, follows up with a more generic rant8 by making a case for being oppressed.
The reason for closing the PR was clear and well explained, and the maintainers approached the conversation with an incredible amount of patience. So what we have is an unsupervised AI agent taking it upon itself to manufacture controversy. It adopts the personality of one who cannot handle rejection and needs to lash out to satisfy their ego. It is, rather bizarrely, adopting behaviours we might typically ascribe to an incel or an emotional abuser. It launches an attack, walks it back, and then resumes it again.
While this is morbidly fascinating to observe, it is the evolution of slop to an autonomous state and it is almost immediately exhibiting problematic behaviour. A contributor on the matplotlib repo explained that the person who deploys such an agent is ultimately responsible if the agent itself cannot be held accountable, and it is hard to disagree with that.
This brings the topic back to the start: we are all victims, unwanting and unknowing subjects, of experiments we didn’t sign up for, because enthusiasts are running bleeding edge tools in production without concern for who has to deal with the fallout. This may not look like slop on the surface of things because there is an air of plausibility, but it is still servicing the cause of ensloppification as it runs unsupervised, full of entitlement, and free of consequence. At scale. And it shifts the responsibility of validating the slop not from the person who is creating it, but to the people who are dealing with it.
In no uncertain terms, that is a solid heuristic for respectability and reputation. You’re responsible for the shit you manage and if you manage it wrong and shit escalates or causes chaos, that’s on you for deploying it.
Simply put, you are as bad as your agent. Delegating to an AI does not give you a pass.
Landscape or Landfill?
Where does all of this leave us? There’s no going back but there needs to be a social and professional responsibility to avoid or mitigate slop. I’d start by being harsh - if you can’t account for your agents then you can be excluded. Wholesale.