Refuting the “weak link” hypothesis

An economist thinks that human “slowness” limits the economy. I disagree.

Like many folks on LinkedIn, I read the Financial Times’ article How much value is AI really creating? with great interest.

It’s undeniable that AI creates digital artifacts, including lines of code, faster than any human could. But humans still need to review that code and evaluate whether it will work as intended. For example, evaluating it for exploitable security holes or inappropriate data disclosure, and reviewing it for excessive resource usage.

Somehow, these necessary, responsible actions—reviewing, evaluating—have been labeled the “weak link” hypothesis. Here’s how the proponent of this, Charles Jones, states it in a working paper for the National Bureau of Economic Research:

“Just as a chain is only as strong as its weakest link, the economy may be limited by whatever tasks are not yet automated — that is, by tasks which are performed by slowly-improving humans rather than by rapidly-improving machines.”

My thesis, illustrated. Screenshot from Insight Fusion’s video about arrows piercing chainmail, with my annotations.

This idea, that humans are the “weak link” that will stifle the (otherwise inevitable) ascendancy of economic growth, is as incorrect as it is insulting.

Let’s analyze this one statement piece by piece, starting at the end.

“rapidly-improving machines.”

I’m not sure that this guy understands what it takes to improve a machine, even an automatic-learning, automatic-adjusting machine. It takes human feedback.

Humans have been the driver of every AI learning from the beginning. It takes humans to define what good is, and it takes millions of pieces of annotated data (some from Captcha, some of it feeding Pokémon Go data to military drones ) for the AI to be able to produce anything.

The major LLMs today were fed content and context from the entire digitized universe, a data set heavily skewed toward American English literary and marketing sources. Then, as people use it, human moderators and data annotators are required to identify and annotate incoming and generated content—a billion-dollar business that depends on the abuse of low-wage workers.

So this change seems rapid, if you limit your scope to only look at the last 5 years. It is incredibly slow, if you consider that this technological theory is more than 50 years old, and even LLMs have been worked on since the 1990s.

“slowly-improving humans”

I’m also not sure that this guy understands what human performance consists of, nor how improvement works. According to our pattern of evolutionary success, humans are the single most intelligent, adaptable species on the planet. We not only seek out patterns, exploit them, and devise new tools that allow us to make impacts at a planetary scale, we have communication and culture that allows a single person’s impact to outlast their own lifetime by a factor of 10 or more.

When it comes to an individual learning to make change, we tend to test and reject changes that are dangerous, rather than run straight at them. We learn incredibly rapidly, while we also seek to preserve and benefit ourselves and our loved ones. I studied this for my masters’ degree in curriculum and instruction; instead of linking every word in this paragraph, here’s a primer on the major theories of human learning.

“the economy may be limited by whatever tasks are not yet automated”

Seriously, this guy is still saying that there’s not enough automation? With an easy, not-even-LLM-powered internet search, you can find the growing evidence that automation itself is limiting the economy.

Even in 2021, before AI, the rise of automation was slowing median wage growth. Automation is was failing to deliver money that the average worker or consumer could spend. That spending by individuals is a key driver for economic growth.

The idea that the economy “may be limited” if any task isn’t automated is ludicrous, and is crucial to his overall thesis. I’m not buying it.

Finally: “a chain is only as strong as its weakest link”

Sure, that works on its surface. I have purchased and used chains, and know how they work. But even this statement doesn’t support his thesis.

Chains don’t know what they’re connecting, and can’t tell when it would be better not to be connected. That’s why “quick release” clips and shackles exist, installed as safety measures so that a human (or even an automation set up by a human!) can act when a chain is pulling something down that should stay up.

Image of a quick release link by Scott Lough on Flickr

In case it’s not clear: this guy is wrong.

He’s not just wrong, his central thesis is easily disproved nonsense.

I’ve worked on AI integration, and I wrote about what it takes to integrate generative LLMs inside software that actually helps people in the second edition of Strategic Writing for UX.

But more importantly, I work as an instructor, mentor, and consultant in how to apply UX skills for business impact. I know how rapidly, how flexibly, how powerfully humans are using the new tools they have access to—and how they’re rejecting its dangers as much as they can while working within the pressures of their own contexts.

As Kate Agena wrote, “A model [AI] can produce fluent text all day. It cannot decide, on its own, whether that text is true to the user, consistent across a hundred screens, and worded so a person can trust it. That is rhetorical work. That is judgment.”

I’d go even further: That is skill, which can be rapidly learned and developed by humans over hours, days, and years. Kim, Maya, and I wrote about the skills necessary to create experiences for humans, and how they connect to actual business impacts that raise the economy, in UX Skills for Business Strategy.

Maybe that’s why this guy’s so-called economic theory is so offensive to me. You, and your skills, aren’t the “weakest link,” no matter what this guy wants to believe. Instead, you’re the strongest point, shattering this fabrication of chains.

What did I get wrong, or right, about this guy’s take on humans and AI? Let me know in the comments.