How is AI going to impact jobs that center around knowledge work? Straightforward mechanical jobs are already being automated away. For example, cashiers are being replaced by vision AI that automatically detects which products customers have removed from shelves, and taxi drivers are being replaced with self-driving cars. Knowledge work on the other hand, seems like it’ll likely be re-shaped by AI rather than replaced by it.

Human-in-the-Loop AI vs. Fully-Autonomous AI

In the next 5-10 years, I believe that most knowledge workers will begin to adopt AI through a human-in-the-loop arrangement, where professionals collaborate with AI to do their work. This stands in contrast to my previously-mentioned examples of closed loop, or fully-autonomous AI, which does work without any human intervention. An example of a human-in-the-loop workflow would be a writer collaborating with ChatGPT. Maybe the writer manually creates an initial draft of a story, and then asks ChatGPT to read that draft come up with ideas for alternative endings. The writer might then write up a draft of their favorite suggested ending, ask ChatGPT to edit it for style or brevity, and then incorporate some of those edits into a final draft.

The distinguishing characteristic of the human-in-the-loop arrangement is that the professional’s role becomes more focused on breaking a large task into small, well-defined parts, instructing the AI to handle those, and then analyzing and integrating the AI’s outputs into some larger piece of work.

Human-AI collaboration of this type makes sense because humans and machines possess different and complementary abilities. Machines excel at handling well-defined repetitive tasks, especially those for which they have seen direct or analogous examples before. On the other hand, people are capable of exercising subjective judgment and generating novelty in ways that machines cannot replicate. Even though AI can outperform people in math competitions, the current state of the art AI technology is nowhere near being able to write a Pulitzer Prize-winning novel, and it doesn’t seem like that capability will be achieved in the near future either.

The Prompt Engineering Challenge

The main challenge with using AI for complex knowledge work is the need to give it a detailed specification of what it should do. Many professionals in creative fields say that they figure out what it is that needs to be done by starting to think through and work through an issue themselves. For instance, many writers claim that they only discover their thoughts and how to articulate them during the process of writing, rather than starting with a precise idea beforehand. Without doing this exploratory work, it’s impossible to come up with a sufficiently detailed specification for an AI tool. I suspect that this holds true for many types of knowledge work: it is only through doing the work that you discover what you want the end-result to look like.

Learning to Leverage AI Tools

The future of knowledge work seems like it’ll involve people augmenting their skills with AI rather becoming obsolete. As AI takes on more and more of the small, well-defined tasks in our workflows, it’ll free us to focus our attention on aspects of our work that require creativity and subjective judgment. This shift necessitates a new mindset, where we learn to view AI as a partner or assistant in our workflows.