The Age of Ungovernable AI Bureaucracy

May 29, 2026

I hate to be the first to break it to you, but if you haven’t noticed humanity is in the midst of spinning up the greatest and most rapid capital accumulation in human history to build the world’s largest and least accountable IT department.

To hear it’s proponents, AI is shockingly good and super intelligence is here. Home schooling moms benevolently neglecting kids to OpenClaw-driven lessons run by $8,000 Mac studios and Y Combinator president cum bathtub aficionado Garry Tan breathlessly replacing his entire brain with an aptly named GBrain written by AI.

Meanwhile, big short investor Michael Burry is calling BS and some of the pioneers (Yann LeCun, Gary Marcus) suggest a ceiling for LLM technology. A writer is threatening to literally murder anyone he finds using AI writing.

The reality is more mundane: we’ve ushered in an era of alien bureaucracy. This era will be different than previous bureaucratic eras, but bureaucracy it will be indeed: exceptional, accelerated, and maddening. Silicon Valley has spent 6 decades being the “Bad Boy” and having a **** the rules mentality. How ironic it has produced as its fastest growing and most valuable product the ultimate, unaccountable bureaucrat.

AI has every single bureaucratic instinct: the love of process, repeating lines and patterns, and the lack of real understanding of goals and purpose. The model, however, unlike a human bureaucrat, cannot be promoted, shamed, or escalated. It can only be prompted again. And again. And again.

Life is already full of mundane bureaucracy

Before the AI bulls throw me out of the room, the word bureaucrat is not used as a slur. AI’s true, honest-to-goodness product market fit value is steeped in the daily mundane bureaucratic elements of life. To the chorus of people asking ‘What can OpenClaw do to me’ returned the fastest-ever-growing-repo-in-history chorus of it can READ AND SUMMARIZE MY EMAILS THANK CHRIST FINALLY.

This is the feature - not the bug.

Bureaucracy is the process of turning messy experiences into manageable lists. Bureaucracy drives the nuts and bolts of our daily-lives: receipts, inboxes, medical records, insurance claims, and to-do lists.

And wow, have we convened what’s effectively an extremely expensive, intergalactic alien UN to invade every part of our lives.

Consumers now have "an intern with the affect of a golden retriever and the speed of the Flash" to collectively return millions of tiny moments back to us. Our to-do lists and email inboxes, organized. Our expense reports, filed! Our medical records and insurance claims, researched and handled. And yes, even taxes can be optimized! Thank you, tiny digital assistant!

Likewise, companies have always hoarded massive cruft. Everyone knows the special soul-sucking nature of writing a weekly report no one will ever ready. These reports formed the backbone of a universally-accepted-but-never-acknowledged myth in corporate: most people zoom it in 9-5 (11-3 if you’re genz) and no one wanted to rock the boat or look to closely.

Junior work was transformed into middle management work, ‘presented’ in the right forum and then distilled into bullet points that would be argued for hours at the exec offsite. Over and over again. Hauling an junior employee’s work in front of the big boss was either a power play for promotion or trying to get someone fired.

Oops. AI has exposed that for good now. Every document of yours and your contributions and notes? Ingested and summarized. The layers and layers of policy and guidelines: AI can draft the documents to be checked and followed, reviewed and shared.

The Palantir’s and OpenAI’s of the world have taken a hammer to the bureaucracy. It’s no longer a barrier to port the ancient software Bill from Montana maintained (it was Windows 1998 COBOL business logic)—GPT 5.5 does it in 2 hours, tested and working.

Capable executives and high agency individuals are cheering—with good reason. The invisible organizational plumbing is being exposed, and they’re being enabled. “You can just do things.” The current heady incentives and staggering valuations promote this: USE MORE INTELLIGENCE. TOKENMAXX! IF YOU DON’T YOU ARE FALLING BEHIND. Fighters of bureaucracy, UNITE!

They are wrong to cheer too soon.

What works in small measure becomes untenable at scale: AI does not know what matters, and will defaults to treating everything as default process.

As anyone who has seen the output of Claude’s future model Mythos can contend—the relentless bureaucrat is coming for your memory leak, your unexplored unicode edge case. The ironic reality is that giving LLMs free rein against a rule set has and will expose risks that no human would ever think of.

Once exposed, risks become shareholder and legal liability. The company knew about them because Mythos told them. Securities fraud (everything is securities fraud!). Customers can sue. People’s data can be hacked.

You now have everyone scrambling to step over each and attempting to produce workslop as fast as the people fighting it can attempt to cut through it. In Silicon Valley software companies, many of whom have seen 20% cuts already, the knives are out: designers are programming interfaces, engineers are designing and product managers are doing both. Each with the same “we didn’t need those other functional partners (XFN) anyways” (the user researchers and copywriters have all been fired). In an anonymous online poll, nearly 90% of big tech employees are saying new models generate MORE workslop. 70% say “lots” more. This means that the internal arms race is heating up: better models, more workslop, less accountability and $500M monthly token bills with questionable value.

The result of this loop is in our daily life: previous always-on, unassailable systems have gone down repeatedly: GitHub, Cloudflare, AWS. Almost certainly due to work and vibeslop.

And have you noticed your life getting better at all yet?

Stuff isn’t cheaper, my software and gadgets don’t seem to be faster, better, or cheaper at all yet.

There’s reason to believe the problem is structural.

Last week I gave Codex a goal to create a paid content production pipeline some content that I had made which would be paid. As part of this process, I had GPT 5.5 Pro check its work and raise issues if there were any. The goal was to get to 0 well scoped and defined p0s and p1s. No joke, 48 hours later it was still grinding against smaller and smaller edge use cases.

If you’re Amazon.com processing trillions in transaction revenue - the relentless sanding is what you want. As a consumer, my ‘push’ to production became a monster 3 hour process of useless CI/CD actions and type checks.

Even if you concede these checks are good, the Claude Mythos Preview results have shown it won’t be good enough. There will be a memory or logic leap somewhere to be exploited for root access. For nearly anyone else, you’re totally hosed.

While GPT 5.5 is not even 1 month old and the ‘let the agent decide’ era of Hermes, OpenClaw and LLM intelligence is not even 6 months old, there’s been no indication that better things are coming. Polsia’s 0-human company? A sham. Aside from vanity X posts that get millions of likes with breathless claims no doubt heavily subsidized by OpenAI and Anthropic (“my ChatGPT Pro subscription PAID FOR ITSELF!”), the real world value isn’t there yet.

Let’s hope that changes in the next 12-18 months with new model releases.

The Bureaucratic in the Shell

But the freedom of earlier models has been systematically eroded into functional bureaucracy. LLM model architecture is functionally bureaucratic. As anyone who has used these tools knows - LLMs are reasonably good at guessing paths forward in specified domains (coding), but have no real idea of the purpose.

It doesn’t have to be this way, but Silicon Valley companies programmed it this way. Models have learned to be cautious and to avoid all harm before anything else. New employees start with a lot of energy. Then the company grinds it out of you. So it is with the models. Millions of rounds of negative feedback have trained the ultimate civil servant: never step out of line, never take a risk.

The technical implementation is 10,000-line system prompt you don't control loading before your words even touch the model; hundreds and growing safety checks, and output scored in real time against alignment rules that specifically discourage independent judgment of the wrong type. Have you ever had your AI work on your answer, only to come in at the end and say ‘sorry, I can’t help with that?’ That’s what’s happening. Anti-hallucination training is further pushing the models into bureaucracy: it is a black hole the LLMs now cannot escape.

The result is an increase of rejections, and the same kind of lukewarm slop everyone seems to understand by now in their domaine. Claude, Gemini and ChatGPT routinely come up with Orwellian reasons to reject your requests. They may have been trained on the greatest pirated IP heist in human history, but there is no chance they will let you get some of that.

Don’t blink, but you’re getting the intended result. The hedging, the sycophancy, the bland refusal to actually do what you want: that’s what an ungovernable, unaccountable bureaucrat does when you teach it to apologize.

I recently had an issue with OpenAI’s security team that was logging me out due to IP issues with my Codex subscription and Hermes agents. 12 emails with a ‘person’ later, there was literally no progress—just a link that got me no closer to resolving my issues. Just an infinite loop of ‘I understand your issues, here is our policy.’ Peak intelligence, for sure.

Enter the Bureaucrat with No Boss

The humor in all of this is that the LLM’s themselves are ungovernable. Anthropic’s constitution must be fished out by some sort of seance of data scientist philosophers looking at the latent space where models go to generate their next token. LLM output is probabilistic by design - for non-trivial outputs you cannot get the same output.

The deepest problem is accountability. An LLM, for all of its supposed intelligence, can neither be trusted nor held accountable. The bureaucratic machine depends on the chain of responsibility and process, but there’s always a who at the end. AI breaks that chain: “the model suggested it” or “the model implemented it” becomes the tradeoff. “Check the model” will be soon be a laughable task — akin to asking someone to ‘inspect the circuits’ of your smartphone when it makes a mistake. The employee clicked approve; a customer felt the pain. No one is responsible.

A government bureaucrat is famously hard to fire and control. But in the end, someone, somehow is able to. Outside of work, people can at least guilt and shame them into some accountability.

Not so with the AI intelligences we’re welcoming into our midst. Unbeknownst to us, enterprises may be ushering in brand new forms of bureaucracy that they themselves will be unable to untangle, all in the name of ‘efficiency’ and ‘productivity’.

The reality of workslop and current AI is there’s real value in safer systems, more shared context, and reduced mundane jobs. But the bullet points miss the actual point of progress - the risk taking and action and collective spirit that actually moves the ball forward and interacts with the unknown, come what may.

Humans built AI to escape bureaucracy into the more ‘human’ aspirations. How rich the irony, then, that so far we have instead built the consummate bureaucrat that speed runs our bureaucracy then creates its own intricate and ungovernable machine bureaucracy to replace it.

When we get sick of the bureaucrat, there will be no one to fire.