Surviving the AI apocalypse is about knowing where to tap

How AI increases the need for human expertise

There is a perception that the impact of AI on knowledge work will be pretty dreary for us.

AI makes expertise more available than ever before. It makes it more accessible (you can just prompt LLMs using natural language), as well as cheaper and faster to produce than humans are capable of.

This means that the production of stuff inevitably increases. Whether you want to churn out marketing copy, summarise reports or even build whole new apps, AI systems enable you to do all of this.

The gap between top-tier experts and average workers has closed dramatically. More people now have access to the knowledge and information traditionally held by experts who would have acquired it via exclusive institutions, industries or other gatekeepers that were essential intermediaries for such resources.

Some people look at this picture and feel an impending sense of doom as an elaborate amalgamation of well-organised sand eats away at the work thought to be human-exclusive endeavours.

I remember when, before LLMs and ChatGPT, AI was seen as a technology that would automate the boring stuff, freeing people to do more creative, fun or intellectually-stimulating work. It was seen as a technology to liberate knowledge workers.

We see much less of this now. Instead what we get is Dario Amodei proclaiming that AI will take half of entry-level white-collar jobs within five years and people losing their minds over it.

Admittedly, there is something strange about the CEO warning us that the technology his company is building will cause this wave of job losses, as if we are supposed to be feebly grateful for the heads up

But this aside, I have been using AI more and more for my work over the past year or so and realised something interesting.

Yes, AI is taking away some of the more menial and boring aspects of my work. I am glad that I do not need to spend time filling out GDPR vendor review templates for example - this is not exactly the pinnacle of my job.

Yet, I do not feel this looming dark cloud that AI will end up just replacing me altogether as the list of things that humans are needed for rapidly shrinks.

I am actually starting to feel the opposite trend happening; AI is giving me more work to do, not less.

So this is what I actually think is happening.

Just because the production of content increases does not mean that the quality of that content also increases. In fact, the quality often decreases and this likely has much to do with how people perceive and use today’s AI systems.

Many people think they are working with some sort of god-in-a-box that you can one-line prompt and get exactly what you need the first time with no mistakes and no iterations required.

This is not the reality at all.

The AI systems we have now are incredibly brittle and path-dependent. When faced with any out-of-distribution task, or given too little instruction or context to work with, they don’t tend to work that well. It takes quite a bit of coaxing, monitoring, nudging and iterative testing and troubleshooting to get AI workflows actually working and flowing.

Verifiable domains like coding are much easier to use with AI - you just run the AI-generated code to see if it works as intended and use this signal to train the models to produce better code. But not all domains are like this. Verifiability is often much more difficult, where what is considered “good” is broader, more subjective and far less certain.

Knowledge work is littered with these hard-to-verify tasks. And it is for this reason that humans are still indispensable parts of the process.

This all reminds me of the ‘knowing where to tap’ story and the value that knowledge workers actually bring to society. It goes something like this:

An old man’s boiler broke down in the middle of winter, so he called a plumber to fix it. After the plumber arrived, he looked at the boiler for a few minutes, took his hammer from his bag, and tapped the boiler on a particular spot. The boiler then started working again, and the plumber handed the old man the bill. “But all you did was hit it with a hammer,” argued the old man. “I could have so easily done that myself.”

“Indeed,” the plumber replied calmly, “but you are not paying me to simply tap the boiler. You are paying me to know where and how to hit it.”

AI systems are hammers, and this is true both in how they work (path-dependent machines heavily reliant on utilising the patterns and distributions garnered from its training data) and how they should really be seen. They are tools that help you execute tasks, and in this way they are brilliant tools capable of doing many different things.

But when you open ChatGPT, Claude or your AI system of choice, it does not tell you what to do nor how to do it. You just have a blank text box with a blinking cursor, patiently waiting for your input.

This is the crucial point. If you see AI as just a tool, then you also start to see all the questions these tools leave unanswered and therefore the things they cannot do for you.

Even in the age of AI, humans are still needed for (a) providing the initial instructions, structure and context, (b) verifying the output and (c) doing something with it.

Without this, blindly trusting the outputs of AI systems used poorly leads to bad decision-making and resource misallocation, both exacerbated by feedback loops and accountability gaps.

This therefore makes human expertise more important than ever. The provisioning of the initial instructions, structure and context, as well as the verification and implementation of outputs, all constitutes knowing where to tap. It is about knowing what to do with the power that AI gives you.

Human experts are the navigators in a rising sea of AI slop. You still definitely apply in the age of AI.