The Death of Bullshit Jobs

The AI versus jobs discussion is mainly about the big (macro) economic implications. Will we see a big rise in joblessness? Or will we see new roles arise? Last week The Economist put "Prepare for an AI jobs apocalypse" on its cover. De Tijd ran its own version: "Hoeveel laat AI straks over van uw job?".

The conversation is hot, but I am loath to make predictions on what will happen to such a complex large scale system. Nor do I have an informed opinion on what kind of new taxation should be designed to fix the resulting economic fabric.

This is a more personal take, what do we see changing on the micro-level? How is the nature of jobs changing right now? What can we reasonably expect in the near future?

Let's cut the crap, and start with bullshit.

Graeber, on bullshit

David Graeber's Bullshit Jobs: A Theory (Simon & Schuster, 2018) grew out of a short essay he published in STRIKE! magazine in 2013, "On the Phenomenon of Bullshit Jobs." His definition is precise.

"a form of paid employment that is so completely pointless, unnecessary, or pernicious that even the employee cannot justify its existence even though, as part of the conditions of employment, the employee feels obliged to pretend that this is not the case"

It is not that this job is unpleasant or hard work. It is a type of work that has no real good reason for existing. Distinct from this is a second phenomenon: bullshit as a byproduct of an otherwise useful job. Generative AI is excellent at producing good-looking bullshit. Long polished documents that look impressive and contain nothing. Slide decks engineered to fill a slot. Status updates rewritten three times before sending. This is the bullshit layer wrapped around real work, and if AI takes that layer off your week, you are unambiguously a winner. Few people got into their profession because they loved the layer.

But back to Graeber's actual interest - could AI be an instrument to get rid of these bullshit jobs? He sorted entirely-bullshit jobs into five types:

• Flunkies: exist to make someone else feel important. The doorman, the elevator operator (post introduction of the elevator buttons), the personal assistant for an executive who does not need one.
• Goons: exist only because competitors employ goons. Telemarketers, much of the more aggressive end of corporate law,...
• Duct tapers: exist because something is broken upstream and nobody fixes it. The person who copy-pastes between two systems that should talk to each other but do not.
• Box tickers: produce the appearance that something useful is happening. Compliance theatre. Reports written for nobody. Meetings whose purpose is to allow a meeting to be said to have taken place. A large share of what currently slows down infrastructure projects in Europe and the United States falls here.
• Taskmasters: middle management. People whose role is to dole out tasks to people who, if left alone, would mostly figure it out.

Which of these jobs are most likely to be impacted by AI?

Duct tapers go first; if a system can talk to another system, the human copy-paste layer between them is the easiest thing in the world to remove. Box tickers shrink because writing a report nobody reads becomes free, so the reports stop justifying themselves. Goons mostly survive, because they exist for symmetric arms-race reasons. If everyone automates the arms race, the arms race continues in a more expensive form.

The death of the middle manager

Perhaps the eroding role of the taskmaster is the most interesting. It is typically the most prevalent and arguably respected role that falls under the category of bullshit jobs. Most people in mid-career end up promoted into one - and somehow end up with much less interesting jobs, even if there is higher pay. Many of the best people I know miss the time when they were making something. Promoted out of the work they were good at, they spend their days routing tasks to others, sitting in meetings about meetings, and writing the same status update to three different audiences.

If agentic AI does what it currently looks like it will do, this layer thins. Not because middle managers get fired in droves, but because the people in those roles can execute again. The senior expert who used to oversee five people doing the work can now do the work directly, with the AI handling the mechanical part at high volume. The coordination layer that existed to manage those five people gets smaller. The person at the top of it stops being a director of others and becomes ten times more productive themselves.

The second-order effect is probably the most universally welcome consequence of the entire AI wave: fewer meetings.

If the coordination layer shrinks, the meetings shrink with it. Death by meeting, the single most reliable destroyer of an otherwise good week, may finally have an antidote. If even half of that promise comes true, it would still make a lot of people happy.

I hedge here for a reason. Plenty of organizations will respond to AI by adding more middle managers, not fewer, because the box-ticking instinct is strong. But the option is on the table in a way it has not been before.

What the Crunch team actually feels

This is where I stop speculating and look at what is in front of me. At Crunch we ran a quick informal survey across people whose work has a technical component: programming, modelling, data work. About 44% say their job has become more rewarding since generative AI showed up. 33% say no change. 22% say it has become slightly less interesting. The tendency is positive but not unanimous, and the people in the 22% are not wrong to feel what they feel.

What separates the two groups, as far as I can tell, is which part of the work they liked best. Software engineering and quantitative work used to be dominated by writing code. The day now leans much more toward design. What is the right question to ask, what is the actual user story, how do these systems fit together, how is this going to be deployed and kept alive. The line-by-line production of code is a much smaller share of the week than it was two years ago. If design is what you love, your job got better. If you loved the typing, less so.

The risk of the Chinese room

There is a darker version of all this, worth naming. John Searle proposed the Chinese room thought experiment in 1980 ("Minds, Brains, and Programs", Behavioral and Brain Sciences). A person sits in a room with a rulebook for translating English into Chinese. English text comes in through a slot, the person follows the rules step by step, Chinese text goes out. The person speaks neither language; the room as a whole appears to. Searle's original target was machine understanding, but the thought experiment now reads more sharply as a warning about the human inside.

The long philosophical fight between Searle and Daniel Dennett is about whether the system in aggregate understands anything. Both sides are worth reading. The more useful observation, for the question I care about here, is the one both camps agree on. The person in the room, the human shuffling paper according to the rulebook, does not understand. That part is not contested. Whatever you think about the system, the operator gets nothing.

A common defense of heavy AI use is that this is simply another step up the abstraction ladder. We went from machine code to assembly, from assembly to C, from C to Python, from Python to natural language; the room is just the next jump. I do not buy it. The previous jumps preserved a chain of understanding. An experienced Python developer can, if pushed, reason about what is happening underneath. The current jump cuts the chain. The simpler explanation is laziness, not climbing. The more competent we make the black box, the less anyone bothers to look inside.

The risk of the next few years is not that AI takes our jobs. It is that we end up as the person in the room. Our day becomes the routing of instructions to agents that actually do the understanding, while we hold the rulebook and shuffle paper through the slot. The understanding has moved out of us and into the system. We are the cog that passes the message along.

That is the failure mode worth watching. The 44% in our survey are not in the room; they have climbed the design ladder, not slid down it. The 22% may be feeling the first hints of the room, and that is worth paying attention to.

So far, what I see is most people staying in their job and finding the work more interesting, not less. That is the better of the two outcomes available, and it is not the one most of the headlines are warning about. Most of us are moving up the ladder, not into the room. Let us keep it that way.