One of my favorite words isn’t found in the Oxford English Dictionary. It hardly has a full page of Google search results. The first written mention of it that I could find is in a Freakonomics podcast episode from seven years ago.

In that episode, economist Stephen Dubner shares a story he heard from a former Dean of Students at Dartmouth College. Dean so-and-so, often caught in fleeting conversations without context, when a response was required in passing but the topic was unclear, resorted to a particular nonsense word.

“Reebusacassafram.”

It doesn’t mean anything. It’s a collection of random syllables, a nonsense word, that sounds like nothing in particular. It's wonderfully nondescript, which makes it a great verbal Rorschach test. People hear reebusacassafram in passing as, “really busy but yes” or “right you are friend” or “great to see you again.” Which is to say, the brains of people listening take something nonsensical and make meaning from it, applying the context of the conversation and projecting the response they want to hear.

Of course, you can’t get away with using reebusacassafram in the middle of an engrossing dinner date. Or while giving a graduation speech. Because, well, context matters. But in the context of passing small talk, where scrutiny is low and the majority of what is said are predictable pleasantries, reebusacassafram (allegedly) works great.

I say ‘allegedly’ because the above story is, well, a story. And that’s okay. It’s the phenomenon reebusacassafram embodies that intrigues me. The hearing-what-you-want-to-hear. It’s pareidolia: the perception of patterns in randomness. Have you ever noticed how the fronts of cars tend to look like faces? That’s pareidolia. Perceiving the pattern of a face where there is no face.

Pareidolia is a nifty manifestation of “representational realism,” the philosophical stance that we don’t perceive the world as it is directly but rather our minds construct an interpretation of the world based on our (very fungible) senses. To a representational realist, there are many versions of our world: the one I see, you see, your dog sees, etc. Each of us has our own “reality tunnels”—mediated by our senses, among other things—that filter in and out the information that constitutes our perception of the world.

In the dawn of artificial intelligence (AI), acknowledging these individual realities matters now more than ever.

II.

I spend a lot of time thinking—and occasionally writing—about AI. At the moment, this means generative AI, a class of artificial intelligence that can generate new content or data, such as text, images, or music, based on learned patterns from a training dataset.

Part of my interest in generative AI stems from the fact I studied creative writing and human-centered design in college, pursuits that at the time I wouldn’t have thought to call “fundamentally human” because the very idea of dividing all tasks into human tasks or AI tasks wasn’t how I perceived the world in 2019. Back then, almost everything was a human task. To the extent I interacted with anything approaching AI, it was in the form of recommendation algorithms, and those always felt comprehensible–as in, you sort of knew how they reached their conclusions (e.g., if user views X movie of Y genre, then show more movies of Y genre to user).

I first became interested in generative AI in 2021. This was in the context of computer-assisted design, and then shortly thereafter, GPT3. Seeing software that could write, and not just write the same thing over and over, or predict my next word (though large language models enabling generative AI do sort of function in this way) but really write something that felt original—that hooked me.

I researched and wrote about generative AI in 2021 and 2022, and then I watched with the rest of the world in November of that year when generative AI went mainstream, with the launch of ChatGPT, and quickly expanded to early-stage commercial applications of all kinds, from generating text and images to designing things you wouldn’t expect, like novel molecular configurations. Progress has been rapid.

Today, AI is capable of generating visual and auditory experiences. But tomorrow, it’s not hard to imagine a world where AI can also replicate smells, tastes, and tactile sensations. Imagine a computer that can interact with you through these senses. It probably feels like science fiction, right? Currently, most human-computer interactions involve basic actions like clicking, scrolling, and typing. And yet, we are moving towards a paradigm where computers understand and communicate in our language, which (1) creates a more intuitive and empathetic interface between humans and machines and (2) makes it possible for more people to interact meaningfully with computers.

As an aside, this evolution underlines my optimism about affective computing, the study of technology that recognizes, interprets, and responds to human emotions and physiological states. Such advancements promise a more human-centric experience of technology, enhancing the ways we exchange information and interact with AI and information technologies generally.

But, back to generative AI. One of the things that makes it so powerful is its ability to tailor content to the individual, to reduce the marginal cost of personalization to near zero. Why is this so extraordinary? Well, for much of human history, people have been unlocking access to greater and greater sources of energy. We used to carry and till and climb and push and do all that by hand. Then we had horses, and therefore one or more horsepower. Fast forward a bit to fossil fuels and we can now haul far more than the equivalent of one human’s daily energy output 10,000 years ago home from Costco in 20 minutes.

But, amidst all that progress, personalization required human brains. Even the assembly line, with all the efficiency it unlocked, made homogenous stuff. To make things unique meant someone had to spend the time and energy to make the thing unique. And, because of this, the cost of doing so could only be so low. A world in which we can personalize so rapidly for so many so cheaply is a fundamentally different world. It’s one in which every student could have an AI tutor, every employee an AI research assistant, every business an AI auditor. You get the idea.

Let’s call this phenomenon “hyper-personalization.” Does it sound appealing to you? In some ways, it appeals to me deeply. In other ways, I am concerned about it. Why? Because hyper-personalization has the potential to be dualistic. As AI tailors content to our personal narratives, it also becomes a reebusacassafram onto which we project our perceptions, and, crucially, our biases. This covers all of the domains in which we interact with AI, as well as our meta-experience of it. In other words, because our general optimism or pessimism for a technology shapes how we interact with it, and because this technology (generative AI) is hyper-attuned to amplify minor differences between us, we can end up in very different places.

What’s worse, this process has a built-in feedback loop: AI's output is influenced by our input, which is then modified by AI's altered output, and so on. Imagine a world where, left unchecked, this process means we end up with two kinds of people: Those who are pro-AI, use it enough to reap some benefits, and continue to do so until they have an entrenched advantage. And those who are anti-AI, refuse to use it, and fall further and further behind, thus embittering them to the technology.

Not only could we end up with a pronounced ability gap, but also we may fail to comprehend the different subjective reality of other people. Without thoughtful, systems-level regulation, interactions with AI may entrench us in our reality tunnels, building thicker walls and creating a feedback loop that reflects and reshapes our worldview through the lens of personalized mediation.

III.

Gertrude Stein once wrote that “everybody gets so much information all day long that they lose their common sense” Put another way, the exponential growth of not only information but hyper-personalized information stands in stark contrast to our fundamentally unchanging capacity to pay attention. One could argue that this disparity leads, for some, to a state of near perpetual distraction. At the very least, our mental space for boredom and contemplation, and, consequently, for creativity, is on the decline. Worryingly, this phenomenon seems to apply particularly to young people, who, amidst an avalanche of digital stimuli and screen time, show a marked decline in imaginative play (something which I’ve written about before) and a worrying increase in myopia, or nearsightedness. While this isn’t entirely because of screens–the bigger driver is the increase in near-to-eye tasks, or, “near work,” of which screens are a large part–it’s estimated that over 50% of the global population will be near-sighted by 2050.

Let’s go back to that Stein quote. In it, she does not claim information erodes our mental space for boredom, contemplation, or even creativity, as I have above. No, she argues it erodes something even more fundamental: our common sense!

Over time, Stein’s insight has only become more relevant. Increasingly, we outsource decision-making to external factors–mainly algorithms and online rating systems, or a combination of the two. Thus, like a muscle weakening due to disuse, our ability to reason from first principles in the absence of external inputs seems, to me, to be on the decline. What takes its place? In many ways, passivity. More specifically: passive consumption of content (e.g., scrolling), and the resulting fee extracted from us in exchange for this passivity (be that in the form of subscription costs, ad-induced spending, or the plain old degradation of our body, mind, and spirit).

But, and I want to be very clear about this, I think on the whole, AI is a net positive. Like any major technology, AI has bad applications, as well as unintended consequences. The trick is to shine a light on as many of those consequences as possible, and set up incentives in a way that we don’t inadvertently encourage individuals or organizations to pursue them (e.g., the profit motive underpinning the capture and sale of our attention by social and traditional media companies).

How do we do this? Well, it’s hard. But let me propose a simple starting point in the form of a framework. It’s really basic. Let’s start with two categories:

Human tasks
AI tasks.

And consider that all tasks fall into one of the two categories—if something is a mixture of both human and AI, as many things today are, then for the sake of this framework whichever entity, human or AI does the majority of work, that is the category to which the task is assigned.

Under this framework, we see that in 2024, far more things can fall into the AI task category than could, in say, October 2022, just prior to the release of ChatGPT. And, I would argue, there are a lot of people spending time today figuring out how to move tasks from the human category to the AI category.

Clear enough, right? Well, it’s important to consider how exactly we are deciding what goes into the AI category. Today, as we explore the many applications of things like large language models, I believe we are in a process of boundary searching, trying to ascertain how big the AI category is. What can AI do? How well can it do it? Students are doing this. Employees are doing this. Chief AI Officers are doing this.

We will need a system to sort tasks into each category. There are and will continue to be sophisticated frameworks for doing this. But for now, let me propose a simple one: subdivide the AI category into “can and should automate it” and “can but shouldn’t automate it.”

To determine which category a tasks falls in, we should apply some basic principles, for instance:

Does automating this save time/money, and/or lead to more effective outcomes?
Does automating this free up human-hours to do something more enriching?
Does automating this result in more equitable outcomes?
If we do automate this, can we easily reverse said automation if we discover undesirable outcomes?

Under these principles, I may choose to automate the writing of some memos or emails, in part because for these forms of writing, standardization is a desirable quality, and in part because I save time, and can then spend that time-savings writing more creative things. And, like riding a bike again after a hiatus, I can pretty easily go back to writing memos and emails.

However, I would not choose to automate the writing of, say, this essay. I may use AI to look up information, or help me brainstorm ideas, or critique my work, but the actual act of writing it—the dedicated thinking through discomfort to create something original—I won’t automate. The process of writing enriches me. There isn’t a higher form of writing for me to pursue. Sure, I could probably do something of higher value, but to me, this form of writing is a form of self-actualization, and so it’s pretty close to the top of my list for how to spend my time. Plus, writing is, to an extent, a skill that must be practiced to be maintained. Outsourcing threatens that practice, and therefore the maintenance of that skill.

We need to begin thinking about intelligence tasks the same way many of us in developed countries think about physical tasks. Though we may have the option to avoid physical discomfort most of the time by driving to work, sitting in temperature-controlled office, coming home and eating on a soft sofa before sleeping on an equally soft mattress—many people choose to seek physical discomfort because it is good for their health. We know the value of going for a walk around the block even if we do not need to do so to forage for dinner. We choose to use our bodies for the sake of bettering our bodies, not necessarily as a means to an end.

We should begin to think of intelligence in the same way.

IV.

Wendell Berry, the American poet, essayist, novelist, and environmentalist, has written extensively about the deleterious effects of mediation between people and the land. On the topic of farming, he writes that people had, historically, communed directly with the land, on foot and by hand, and in doing so, had learned from and nurtured it, at a sustainable scale.

However, with the rise of agribusiness and the advent of mechanized agricultural equipment like the tractor, we’ve been able to grow more calories than ever before, but, as Berry points out, farming is no longer about interacting personally with the land. Instead, it is about interacting with a machine that in turn interacts with the land on our behalf.

This is to say, the farmer sits in his tractor interfacing with the dirt. From his perch, it is difficult for the farmer to observe the slight discoloration on the leaf of a plant that could indicate a malady, or to sense soil dryness. To compensate, the farmer may deploy apparati of sensing—of data-collection—which ultimately may boost crop yields even more. In this way, the farmer builds up the machine infrastructure that in turn offloads the necessity of paying attention. The abundance of information erodes and supplants what had been common sense. And, in the process of building up this interface (and in the absence of effective regulation), the farmer becomes more and more indebted to the providers of the equipment. Providers that have never and will never interact directly with this specific tract of land. Through this process, the local farmer becomes dependent, detached, poorer, and ultimately, replaced—his land bought up and combined with other newly purchases plots to perform agriculture at a greater scale.

Technologically-induced mediation does not always end this way. In fact, increased technologically-induced mediation is a prominent trend throughout human history, and one which, often times, has been good—basic tools like the lever, pulley, and wheel mediated our direct interfacing with rocks, buckets of water, and cart paths, which led to additional time and resources, which drove technological progress, which overall has been net positive. But that does not give us license to ignore the positive and negatives that are masked by such a label as ‘net positive.’

Furthermore, even if such an effect is net positive, it is often the case that some people are wary of the technologies that perform mediation, especially when such technologies are information technologies.

One needn’t look further than Plato’s Phaedrus to understand that some of the best minds of their day worried about the advent of writing and how it might weaken our mental faculties, which had hitherto been sharpened through the practice of rhetoric.

Similarly, in the mid-1800s, the rise of train travel coincided with the proliferation of inexpensive paperback books. During this time, some intellectuals, such as British geologist Charles Lyell, bemoaned that such small print would damage the eyesight of readers (it probably did not, but the general trend towards near-work has), while other intellectuals noted how antisocial passengers had become: engrossed in their books, as opposed to interacting with their fellow passengers. The locomotive itself compounded this shift: compared to horse-drawn carriage travel, which took three times as long on average and consisted of sitting close to one’s fellow passengers, train travel was shorter, less social, and more impersonal.

All of this is to say it’s not uncommon for new technologies to instill a moral panic about what will happen as they render prior forms less important. Generative AI is no different. Neither all good nor all bad. Inspiring both excitement and apprehension.

Those initial conditions—excitement and apprehension—matter: They interact with and are amplified by other factors like internet access, household income, job type, and more. And taken together, they can be enough to set us careening down very different reality tunnels.

I don’t have the answer. But through the process of writing this essay, I’ve arrived at a good starting point, a question.

How should we choose to organize our lives, and furthermore, our society when it is possible now more than ever to mediate our sensory experiences of the world?