From Cory Doctorow’s indispensable Pluralistic:
That’s the true cost of all the automation-driven unemployment criti-hype: while we’re nowhere near a place where bots can steal your job, we’re certainly at the point where your boss can be suckered into firing you and replacing you with a bot that fails at doing your job
I pay my bills designing and building AIs — I’m professionally (and attitudinally) predisposed in favor of them. But both a theoretical understanding of how AIs work and pretty much everybody’s experience with them show that we don’t really have jobs that map well to the capabilities in what we currently call “AI” (one exception: the sort of at-scale meaning-free bullshit writing that some people are paid pennies for but that we don’t really want more of, do we?).
There are three rough categories of business AI projects:
- Replace people. The vast majority of projects. The first 80% gets done very quickly, the last 20% never: mindlessly extending text maps better to some managers’ view of what professionals do than to what they actually do or should do. Projects stall, die, or are declared a success and then the PR and legal teams have to deal with the fallout of deploying them.
- Make people more productive. A growing minority of projects. They often work! Generating plausible text, code drafts, etc, is part of most people’s professional activities, and AIs (again, in the early 2020s’ colloquial use of the term) are very good at that. The side effect to be mindful of is that this is part of most people’s activities, and sometimes it’s the most visible part, but it’s seldom how they add value. The risk is that making writing slides or code faster, but not doing the thinking that goes into it, can lead to more slides or code but with less thinking.
- Make new things possible. A tiny minority of projects: by definition, very few companies are at the cutting edge of their business or make being there their core strategic bet. These are the high-risk, transformative, redefine-the-market-it-works projects where computational intelligence beyond last year’s state of the art can lead to business capabilities beyond last year’s state of the art. This doesn’t need to “look like AI,” only to be computationally smarter than what existed before: Google was built, after all, on an algorithm that was highly technical, philosophically unremarkable, and orders of magnitude beyond what existed before.
As a rule of thumb, the first type of project is a dead end for everybody, the second type of project is a good fit for large organizations where large improvements in the productivity of specific tasks can have a relevant aggregate impact, and the third type of project is at the core of high-risk/high-upside early stage startups (or analogous projects inside larger organizations).