We Are (Probably) Not All Doomed

Because of the work I do, I am online a lot. I don’t know if that makes me, as some describe it, “extremely online,” but I am certainly dialed into Internet discourse to a significant enough degree that I am exposed to most of the discussions and disputes firing up social media.

 

One of those topics is AI Doomerism—the idea that the expanded improvement and use of artificial intelligence is leading us to a dystopian future where we will all be unemployed, or AI-controlled robots will enslave us, or that we are headed inexorably toward something called The Singularity, in which artificial superintelligence causes the extinction of the human race (or maybe just locks us all out of our DoorDash accounts; I’m admittedly a little fuzzy on the particulars).

I share some of these concerns, fueled by what some companies are branding as AI-created layoffs at many companies as algorithms replace some analysts, programmers, and administrative staff. AI very well could lead to significant job losses in industries where desk workers do a lot of research, analysis, and report writing. It is also possible, of course, that the current round of layoffs are being “AI-washed” to conceal overhiring or poor financial performance. I think it’s also possible that poorly informed executives may cut a lot of staff because they assume AI can easily and cheaply take on those tasks, and that those executives may be in for a rude awakening.

The key reason I have not sunk into hopeless AI Doomerism is that I have talked to a lot of people who have developed these AI models or use them for specific engineering tasks. Large language models (LLMs), which are the type of AI model most of us are familiar with, are still vulnerable to hallucination and model collapse, and while the models keep getting better, those two problems are not entirely solvable.

In fact, according to AI engineers I’ve spoken to (who surely know a lot more than I do), the problems of hallucination/collapse are only manageable through a significant amount of data curation. That requires people, and it also requires good data based on actual tests, real-world inputs, experimentation, and expert analysis. Even synthetic data requires a lot of management and gate keeping.

Without that human-in-the-loop, LLMs may provide incorrect information or even forget how to do the tasks they were designed to perform. Of course, I’m not sure which idea is more disturbing: artificial superintelligence that will upend civilization, or the possibility that captains of industry will be led to catastrophe by AI models suffering from an acute form of digital dementia.

From what I’ve seen in the design and engineering space, AI looks to be an interesting tool for increasing productivity, and could possibly be a revolutionary tool for improving design quality. AI will probably change the way we work and live in significant ways in general, but it may also fail to do so in a lot of specific instances. In other words, we face a lot of uncertainty when it comes to AI; fortunately, engineers and scientists are pretty good at grappling with uncertainty.