How AI work changes us

Over time, doing work may increasingly mean doing work with AI, but what AI work does to us is still unclear. Some early research points to increased productivity alongside more worrying trends, including disengagement, cognitive offloading, and the risk of de-skilling more generally.¹

AI is not wildly unique in shaping us. The infamous London taxi driver study, which brain-imaged prospective cabbies over 3–4 years as they learned the London streets in preparation for testing, found increased grey matter in a part of the brain dealing with spatial memory among those who qualified.² On the other hand, GPS users have been shown to learn routes less well,³ similar to the “Google” effect where people remember less of information they expect to be able to access later, only a web-search (or, these days, an AI ask) away.⁴

At the same time, AI is unique, at least compared to other technologies, in several important ways that have implications for how we work with it.

Responsiveness

AI is getting better and better at giving us what we ask for. For any question, it can scale the depth, complexity, tone, and style of its answer to whatever the user requests. It can offer, or not, detailed information about what it’s doing or how it plans to approach a task.

But, it’s responsive, not directive, so the user is never truly required to learn the ins and outs of the thing they’re building or deciding on. AI will give us what we ask for, but that neverending willingness means we may only learn and grow as much as we actively try to.

Knowledge

The training data used by frontier labs is legendary in its scope and expanse. Vast amounts of internet data, books, human annotators annotating relentlessly.⁵ And it shows! For frontier models, they can offer deep, even groundbreaking answers to difficult questions.

At the same time, all too classic automation biases complement LLMs’ facility at producing persuasive-sounding sentences to reduce our skepticism of AI outputs and create a reliance on its information and framing.⁶

Capability

AI can just do things, including very complicated, seemingly impossible, groundbreaking things. It wasn’t very long ago at time of writing that OpenAI announced a proposed solution to the Navier–Stokes problem, along with approximately a bajillion other math problems.⁷ Some experienced AI coders say they look at their code much less than they used to, but still produce apparently effective results (e.g., Boris Cherny).⁸

But, AI’s ability to do things means it’s used to do the things that teach us how to work effectively. Researchers have already sounded the alarm about entry-level employees potentially missing the opportunity to learn the skills that would make them highly effective mid-level and senior employees.⁹

Ease of use

Starting work with AI can be as easy as saying what you want done, which, some researchers have noticed, means people can end up using AI to work more and more.¹⁰ AI work bleeds into lunchtime, a quick prompt after work, perhaps an agent ping in the shower. It’s very easy to get started working with AI, helping us get more work done.

But, as in the above examples, that can lead to work bleeding into previously non-work times. AI’s natural language ease-of-use reduces the effort to produce work, potentially Jevons-paradoxing us into an expansion of what we end up doing.

Of course, I write all this as a relatively heavy AI user, somewhat nervous about the tendencies I notice appearing in my own work. As valuable as I find AI as a tool, I am worried about how working with it will change me if I’m not careful. Specifically I worry about:

• Surface-level engagement - I don’t engage deeply with the topics because I just assume AI can do it. I spend a lot of time on different kinds of software development, but I know I’ve barely scratched the surface, and I worry AI may stop me from scratching in the right ways.
• Assumed knowledge - I convince myself I more deeply understand subjects than I do because of convincing AI summaries. Or, I trust the sources AI pulls are the sources I most need to read, that something crucial wasn’t missed.
• Passiveness - I let AI make more and more decisions, accept more and more of its suggestions, slowly let the effort I put into my work recede.
• Focus - I split my time across a bunch of tasks, never learning to carry one thing through to an excellent finish.

So, how do we engage with our work using AI so that we can not only avoid the negative consequences mentioned at the top, but actively cultivate the qualities we want our work to instill? Although I’ll offer some early ideas, generative AI, my understanding of it, and research into its effects are all very early.

How I’ve started to think of it (though, with the eternal grain of salt that I have a lot to learn on this and every topic), is by asking what is the work I’m actually doing and, perhaps more importantly, the work I’m not doing? I ask, because though I worry about passive engagement and overreliance, with or without AI the goal isn’t effort and engagement for its own sake. Highly engaged, extremely drudgerous work is probably a bad use of our time. Having AI take on a large share of those efforts, even if it means our disengagement from them, feels pretty alright.

Beyond that, I’d even go so far as to say there is a degree of de-skilling I’m not too upset about. I’m pretty comfortable weakening my skills at building dashboards in Power BI so long as I’m still building truly helpful data visualizations some other way.

The problem is, then, the moments in a project when I believe my attention is important, and I’m putting it elsewhere. I think about times, especially at the end of the day, when I catch myself saying “go ahead” to an AI plan I’d only skimmed because I didn’t want to go through and give each and every decision its due (and time-consuming) consideration.

That kind of meat-proxy rubber-stamping presumes, implicitly if not full-throatedly, an inability to offer value beyond what the AI has plopped out. Which, given how good AI is at coding, especially compared to me, is not a bad bet. But, my inability to offer perspective is itself a trigger to ask and learn enough to make an informed decision, even if that decision is just “Ok AI, go ahead.”

To change, we need to set up our work so that the things required of us to learn and grow exist within our workday. Those will be relative to what we want and what we’re doing. I’ll share some brief examples of what this has looked like for me.

Making demands of ourselves

Work makes annoying demands of us that also make us more effective. Like learning code by tracing a bug, or improving prose by struggling with a sentence, the problems we’re forced to face just to get to the end of a project are what make us good at the things we care about. Except, if we let it, AI can skip to the end, filling in the blanks, removing the annoying, necessary, eat-your-greens work that we learn from.

But, accepting AI output as-is is our choice. Pushing back on AI suggestions, digging into the underlying topic, taking uncertainty as an opportunity to investigate, even doing the annoying parts by choice, those are the demands we now have to make of ourselves if the shape of our work no longer makes them for us. For example, one study found that users who reported using AI for hints or clarifications showed similar performance and persistence to the control group on a subsequent unaided task, while those who used it for direct answers performed worse and skipped more problems. Those usage patterns were not randomly assigned.¹¹

Embed active review/feedback practice, with etiquette

AI makes it a lot easier to ‘just do things’ which also makes it easier to investigate less and request less feedback than would end up being valuable for the project. Asking for help from others, a sometimes annoying, occasionally time-consuming, highly valuable process, also gave us an opportunity to gather insight and guidance we weren’t explicitly looking for, not as common an occurrence with as deeply responsive a technology as AI.

Even in a world of AI, human review and external feedback matter, so deciding on triggers for deeper investigation can be valuable. There are more or less formal methods, for example some German researchers studying how AI practitioners approach review found 9 practices, including scaling delegation and scrutiny to the task’s stakes, reserving sensitive or critical passages for fully human authorship, asking AI to identify weaknesses to focus on, pausing before accepting polished output, and persistently revising AI wording.¹²

In my case, for example, that involves preemptively deciding what matters enough to actively decide, alone or with others. That includes features, processes, language, etc, with the degree of feedback dependent on the stage of the project (and the willingness of the reviewers). One other caveat: sharing AI outputs is only natural in a world where we work with AI, but etiquette is an ever-present necessity. We have a responsibility to ensure our outputs are clear and digestible, triggers or no.

Regular old learning & coursework still matter

With AI, we can do things we don’t understand all day long, and it can feel really, really good. I live in unending terror of the decisions AI has made that I don’t even know enough to ask about. I can’t count the number of back-end bugs I shot myself in the foot with by failing to understand enough about the topic to ask AI how it works, or the number of UIs I’ve scrapped because they were frustrating to use in ways I can’t articulate.

Thankfully, people have spent a lot of time trying to teach other people how to do things, and all of that material is still around. I spend more time learning coding now than I did before AI, and often from resources that predate genAI’s existence. Using AI to learn can be effective,¹³ but I often worry I don’t know enough to know what I need to find out. As a result, classic educational materials, courses, videos, and organizations like freeCodeCamp offer independent perspective that grounds me in a way I struggle to feel when asking AI directly.