Artificial Intelligence and the Death of the “Power User”

Looking up at that title, it’s feels like a bit much — doesn’t it? Honestly, it does to me — and I wrote it. But I think there’s a truth to it, and it’s one that hasn’t really been brought up much, at least that I’ve seen.

What I want to talk about is, I believe AI (and machine learning in general) have been gradually forcing a fundamental change in the way we interact with the systems we depend on, even as those systems become more increasingly vital aspects of day-to-day life. AI is reconfiguring the way we use our systems, the way those systems serve up their deliverables to us (whether data or something fuzzier), and what it means to be good at using a system.

Specifically, this change affects what it means to be a tool-user: it’s a change both to the types of tools offered (and the modes of use supported by those tools), as well as to how we conceptualize and act on the intent of the person “holding” the tool

It has been said that fundamentally, humans are “tool users”. We identify a situation where we need something better than our own body to accomplish a task, and we refine that something until we’ve produced an article that extends our ability beyond what we were previously capable of.

Before we get in too deep, let’s get one thing out of the way: tools aren’t just physical objects we can hold. Google Search is a tool in the same sense that a club or spear is a tool: they’re constructs, refined to varying degrees, that make us better at something we need to do. In the case of Google Search, what we’re really using is a tool comprised of many smaller, simpler tools — closer conceptually to a pocket watch or a calculator than a spear, really — but a tool all the same.

Historically, there has always been a sort of push-pull between two primary modes of tool use. Maybe it’s better to call them “principles”:

• We make tools to fit our needs, and we refine them based on our growing understanding of what makes a “good” tool . Fundamentally, this is the principle of iterative design, and also touches on the concept of ergonomics.
• We learn how best to use the tools we make, and our skill with a tool grows as we identify how best to conform our actions to those that are most aligned with the tool’s strengths . Or, put more concisely: operational attunement.

It’s from the second principle that the “power user” from the title of this article arises. Power users are users who have a deep understanding of the way a tool operates, both on and beneath the surface. This understanding allows them to use the tool in ways that, while perhaps not intuitive to the uninitiated user, are closely attuned to the tool.

The design and use of every tool pulls a little (or a lot) from both of those points.

That “tools made of tools” concept I mentioned just a few moments ago, in relation to Google Search, is an important one. The vast majority of our modern toolbox consists of these types of “devices”: smaller pieces combined to create something more powerful — with the resulting tool then later incorporated into an even more powerful and useful tool.

In order to be an effective tool user, you need to understand both what the tool is doing when you use it (i.e. how it works) and how its performance changes as you vary your use of it. Someone who is skilled with a particular tool (e.g., our power user) has, in effect, created a complex mental map of inputs and outputs that allows them to produce their desired results without resorting to trial and error. They understand the tool, and using it becomes second-nature.

When a tool is incorporated into another device, our access to the ins and outs of the original tool change. The original tool’s functionality is constrained in some way, restricting its “flexibility” to different modes of use — but we gain consistency and control in other areas. Often, this makes the tool more accessible to a different class of users by lowering the baseline skill required to use it— and this frequently comes at the expense of those who were most highly-skilled with the original tool.

But that’s all vague and theoretical. Let’s consider the real-world example of a common wood-working tool: the chisel.

A chisel is fundamentally a simple machine (an inclined plane). It consists of a handle attached to a metal shaft and blade. The blade has a bevel ground onto it at a specified angle, which is then finely sharpened and honed. Through the use of a chisel, the user is able to remove layers of wood from the surface of a piece for the purpose of smoothing, shaping, or otherwise modifying the wood for use in some sort of end product (e.g. sculpture, furniture, decorative moulding). (You could argue that the chisel itself is a more specialized version of a simpler tool — the knife — and I actually agree; but I think this is as good a place to start as any).

A skilled operator with a practiced hand can use a basic chisel, kept exceptionally sharp, to perform a variety of tasks. One of those tasks is smoothing a piece of wood to a level, uniform flatness. But that task isn’t an easy one; all it takes to damage the piece is a slight slip of the hand, too much pressure applied at the wrong moment, inconsistencies in the wood grain, or a flaw in the tool’s edge. The practiced user can overcome or outright avoid these pitfalls, but the novice will likely encounter them frequently, limiting their use of the tool until they’ve developed sufficient skill and familiarity with it. Because of this, we’d say the chisel has a great deal of flexibility and versatility, but a high barrier to entry.

Now, consider another tool: the hand plane. A hand plane (or smoothing plane, or simply “plane”) is, in essence, the blade of a chisel that has been attached to a solid base. A slot in the bottom allows the blade to protrude by varying degrees, giving the user the ability to control the depth the chisel will cut at. The base of the plane is long and flat, and slides smoothly along the surface of the wood. It acts as a depth stop for the chisel by preventing it from digging more deeply than the amount it protrudes through the base.

Unlike a chisel, a plane doesn’t offer much freedom to turn the tool or alter its cutting behavior during use. You can’t carve with it the way you can a chisel — and you might argue that’s the point. The plane is for smoothing the surface of a wooden board, and it does that very well: the user need only push the tool across the wood at an even pace with consistent downward/forward pressure to shave an even layer off the piece. In creating the plane, we’ve in effect lost some of the features of the original tool, but we’ve strengthened a core mode of use and made it more accessible to the average user.

Then, consider a third tool: the electric plane. An electric plane takes the idea of the hand plane and converts it into an electric power tool. We retain the smooth base and general shape of the prior tool, but we trade the blade for a rapidly-spinning cylinder with a sharpened insert (the “chisel”) that cuts across the surface of the wood thousands of times per minute. Moreover, we’ve traded the simple blade adjustment screw and clamp for an adjustable “shoe” that creates a height differential across the base plate of the tool, exposing only a specified amount of the surface of the wood to the spinning blade.

Where hand planes and chisels are quiet and methodical tools, the electric plane is noisy, aggressive, and powerful. Rather than shaving long, thin curls of wood from the work piece, it cuts a shower of tiny chips from the wood. The rapid speed and power of the spinning blade allows it to cut cleanly and smooth effectively in situations where the grain of the wood might have made hand planing difficult, and chiseling near-to-impossible. Where chisels and planes need to be sharpened and used with care to cut cleanly (and can’t remove large amounts of wood without fracturing the piece), the electric plane can shave an eighth of a inch or more off a piece in a single pass. We’ve lost the finesse and grace of the hand plane, and nearly all of our obvious lineage to the chisel — but we’ve gained a strong, consistent tool that requires much less training in order to use. It would be a massive undertaking to shave an inch off the bottom of a wooden door using only a chisel. It would be possible using a hand plane with a sharp blade and a steady hand. But with the electric plane, virtually anyone can perform the task.

Finally, let’s take a look at our last example tool: the thickness planer. A thickness planer takes the concept of the electric plane (spinning cylinder with a chisel edge, shaving irregularities off the surface of a board) and optimizes it for a single task: shaving a wooden board to perfect flatness on both faces. Instead of a handheld tool, we have a free-standing tabletop machine that uses one or more cylindrical rotating blade to shave the entire surface of the board at once. We’re limited to whatever width and depth board the tool is built to accommodate. The tool is heavy, expensive, and single-purpose. It’s also highly-specialized and delivers a very consistent, high-quality result.

So, let’s bring this back to the thesis of our article for a moment: What’s the point of all the above (other than giving me an opportunity to talk about woodworking tools for a moment)?

It’s about what the evolution of a tool does to its user base (both positive and negative:

• It makes the functionality of the tool accessible to a much wider audience (by lowering the entry level skill required to participate and stabilizing the outcomes that result from using the tool)
• It devalues the accumulated skills built up by users of the original tool (by removing the necessity to accrue that skill simply to perform a given task; this can be seen quite clearly in the shift from film photography to digital photography: previously, simply taking and developing a decent photograph at all required specialist training and expensive gear; all that changed with the commoditization of basic image capture.)
• It raises the baseline quality of output produced by tool users for standardized tasks(because the machine is doing a greater portion of the work)
• It homogenizes expression by ensuring the the majority of tool users are only exposed to the highly-sculpted functionality of stage 2, 3, and 4 tools. Consequently, this narrows their conception of what the tool can do, which shifts user demand away from the flexibility offered by the original stage 1 tool.

That last one is, to me, one of the most interesting bits. I believe the way we interact with the world — our mental model — is defined (and constrained) by the ways we believe things are supposed to work. When we remove the variety and range of expression from a tool, to the average new user (or even non-user, who simply encounters it in the periphery) we’ve removed the concept of performing that action.

If we don’t know something is possible, or how it can be done, we don’t think of ways to do it, and we don’t design our expectations around it. This further suppresses that mode of expression, and makes it even less likely to be afforded by the next iteration of the tool.

When the operation of a tool has been obfuscated so thoroughly that you can use it for a highly-specific purpose and not really understand anything about what the tool is doing, you’ve reduced a previously discoverable skill to magic.

Okay, so say that’s all true; who cares?

Me, for one. And I think you should care, too.

AI is driving search to be anticipatory and fuzzy, interpreting requests to get at what the model “thinks” we wanted, rather than literally what we said we wanted. That’s pretty amazing, when you think about it — and there are definitely times when it’s a (sometimes literal) lifesaver — but it also can (and does) create some very real issues.

The power users of search in years past were able to anticipate the way the algorithm would interpret their query and the content in the index, and use that knowledge to find content that was highly specific. Modern search pushes users to the content that is believed to be most relevant. Not relevant to the user’s query, but rather to the moment/trends/context as well (or more so). This may be more valuable to the collective “relevance” of search as a whole — but that value comes at the loss of entire strata of information we were (perhaps only a decade ago) able to access with ease.

While writing this, I kept coming back to the concept of Newspeak from George Orwell’s 1984 — but there’s a key difference between the two scenarios:

• Newspeak compressed shades of meaning and complex thought into simple vocabulary. This simplification literally removed the people’s ability to express certain thoughts, because there was no way to conceive of them.
• AI-Driven Search is not (at least so far) removing the ability to express thoughts — but it is blurring the lines that traditionally have separated skilled inputs from clumsy ones. That still has the effect of simplifying expression, but the feel of it is different.

Let’s get weirdly specific for a minute: Imagine, if you can, an AI-assisted hammer that you hold in your hand, just like a regular hammer. But hold on a minute; this hammer is smart: it’s able to analyze the hardness of the material you are swinging it at, and to vary the force being applied by each blow in real-time in order to maximize its power without damaging the surface of the workpiece. No matter if you swing hard, soft, clumsily, or with precision — your blow will land with the right force.

Now, imagine how your control over the actual force applied by your swing (if the AI were to be turned off) would be affected by this hammer. Say you’ve swung it for 20 years, slowly building a lifetime of muscle memory that tells you there’s no significant information in the speed with which you move your arm. How accurate would your swing be, then? How good are you really at swinging a hammer? Does it even matter?

Or, to bring this all squarely back into the here and now:

• How much of your skill at texting on your phone’s on-screen keyboard is due to the software’s interpretation of where you meant to press versus where you actually pressed?
• How good are you at spelling when autocorrect and on-screen suggestions are always available?
• Are suggested words and phrases narrowing your usage of language by pushing you deeper into a self-reinforcing feedback loop ?
• Was Google search better when the average user had difficulty finding commonly-sought-after information (but the skilled user could easily find esoteric and obscure information)?
• Or is Google search better now that the average user can quickly find popular content even through a poorly-formed query (but the skilled user is often unable to find anything at all due to the algorithm de-prioritizing unpopular content)?

To all the above: I don’t know.

…but I definitely have more than one type of chisel in my workshop.