I just finished R.F. Kuang's Babel, a novel about a 19th-century Oxford institute where scholars channel meaning through silver bars. Note: light spoiler alert, keeping it high-level enough for the book to still be exciting for you. Pair the right two words across languages, etch them onto a bar, and the bar produces a small piece of magic. The empire runs on it. Carriages move faster, factories produce more, bridges can be built taller and longer, all because of those silver bars and the scholars who craft the match-pairs.
After many non-fiction books I was excited about escaping into a fictional world until I realized just how many parallels there are to my day-job with AI: replace silver with LLM model weights, scholars with prompt engineers and tool builders, and match-pairs with the skills, scaffolds, and context files we now bolt onto large language models, and you get real magic and the risk of over-dependency.
Parallels between fiction and non-fiction kept me thinking back & forth
In Babel, silver unlocks things that simply were not possible before. The capabilities don't ladder up from existing tools; they appear, fully formed, the moment the right pair of words is etched and the words are spoken correctly with full intent. AI feels the same way to me right now. Dashboards I would have spent three weeks on take an afternoon. A messy spreadsheet becomes a structured analysis in a single session. Documentation I would have postponed for months gets drafted, reviewed, and published in days.
And like the scholars in Babel, we are still discovering new match-pairs. Every month brings a new context-management trick, a new agent pattern, a new skill that lets the model do something it couldn't do last quarter. The pace is intoxicating (and addicting).
But here is the part that has stayed with me since closing the book. Take away the silver, and Babel's empire doesn't just slow down. It breaks. The ships leak, the carriages stop, the bridge collapses. The system had quietly absorbed silver into its load-bearing structure, and no one had a plan for what to do without it.
Why this isn't another Industrial Revolution
A common counter is that every technological shift looks scary in the moment and works out fine. The cotton mill, the steam engine, the personal computer. Society adapts, people upskill, we evolve.
I think AI is genuinely different on one specific dimension: where the magic lives. At the risk of over-simplifying: a car is a car, a printing press is a printing press. You can rebuild one in a town with enough wood and iron, and you understand it well enough to fix it when it breaks. The Industrial Revolution distributed both the benefit and the means of production. Even at small scale, it was replicable.
Frontier AI is different. It is powered by proprietary and often black-box parameters and a small number of extraordinarily expensive data centers, owned by a small number of companies. There is no village-scale frontier model (yet). There is no garage-shop foundation model that catches up over a weekend. The magic, like Babel's silver, is concentrated by design.
And the dependency compounds. We are now building systems with AI that are too complex for humans to fully comprehend, on top of older systems that were already too complex for humans to fully comprehend. The stack grows faster than our ability to audit it.
Two failure modes I've already seen myself
Let me make this concrete with two patterns I've watched play out, including in my own work.
The first is expectation creep. People get used to what AI can deliver: the polished dashboard, the synthesized briefing, the comprehensive comparison table. The artifact becomes the new normal. Then someone changes teams, loses access to the better tooling (Claude Code one quarter, plain chat the next), and is suddenly unable to produce the kind of output their colleagues have come to expect from them. The work didn't get harder but the magic just got pulled.
The second is the one-person-doing-three-jobs pattern. Resource-strapped teams quietly redistribute scope onto whoever is still standing, myself included. AI absorbs a meaningful share of those new responsibilities, sometimes well, sometimes barely. It is sustainable as long as the token budget holds, the knowledge inside the agent stays current, and nothing breaks in the underlying tooling. Pull any one of those threads, and the whole arrangement unravels.
I wrote in an earlier post about how AI helped me save 150 hours on a refresh analysis, and how the same workflow nearly produced a 5.6x overstatement. The savings and the risk are two sides of the same coin.
What to do in the short term
If we accept that some version of this dependency is here to stay, three things help, and none of them are exotic.
Keep building your critical thinking. It is genuinely easier to trust the AI than to follow how it got somewhere, and that ease is exactly the problem. Even if a less-powerful model is all you have on the fallback day, you can still get useful work out of it if you understand the structure of what was built.
Document like the silver might run out. The 'boring' engineering disciplines, schema definitions, assumption logs, README files, are more important at AI speed, not less. The good news is that the same tools that increase your output can write a lot of that documentation too.
Run independent QA. If one AI built the thing, a different process, ideally with different people, should re-check it. This is just security and quality engineering applied to a faster pipeline.
If you have the compute capacity: run and experiment with local LLM models. Having your local LLM can be a means of independence, validation, and preserved privacy. Local or not, make sure to build strong guardrails and security.
Zooming out and looking ahead
The longer-term work is societal, not personal. I'll be brief because others are saying it better. Ethical, human-first design needs to be the default rather than the aftermarket option. Access to capable AI needs to broaden, not concentrate. Norms and regulation need to catch up before incidents force them to.
I found it surprising to find Pope Leo XIV and a 19th-century fantasy novel landing on the same metaphor (and the Pope engaging directly in a technological evolution). The Pope's recent encyclical Magnifica Humanitas and his presentation remarks frame AI through the image of Babel directly, and the Wall Street Journal coverage makes the point accessible if the source documents feel dense. You don't have to share his theology to take the warning seriously. The tower in Genesis didn't fall because the bricks were bad. It fell because the people building it lost the ability to understand each other.
That is the risk worth naming. Not that AI fails, but that we keep stacking it on top of itself, faster than we can read what we just built.
For now, the silver still works. The right move is to enjoy the magic, keep learning languages and thinking of match-pairs, and diligently make sure we know how to live without it.