What's the Deal With Machines That Think They Know Things?

There are two ways to make a computer seem smart, and both of them are, frankly, a little quirky when you look closely .

Way one: you find genuinely brilliant human(s), sit them down in a room, and interrogate them for months until you've squeezed their entire career into a very long list of IF-THEN-ELSE statements . Way two: you skip the human entirely, dump the internet into a machine, and let it figure out its own rules, none of which it will explain to you, ever, even under oath .

The first one is an expert system . The second one is a neural network .

And no, we're not getting into AGI or ASI here . That's a story for another day — specifically, a day when literally anyone in the field can agree on what those words even mean .

Exhibit A: TurboTax, the world's most polite bureaucrat

You probably have used an expert system . It's called TurboTax, and it has been quietly interrogating Americans since the 1990s using exactly the architecture Ed Feigenbaum was worried about back in 1983 — a giant nested pile of "if this, then ask that" logic, built by tax experts who had to explain the entire U.S. tax code to a room of engineers, one maddening edge case at a time .

It never learns anything on its own . It never generalizes . If you have a weird enough situation — foreign income, three side hustles, an Alpaca farm — it just keeps asking you gentle, increasingly desperate questions until you give up and call a CPA .

This is not a bug . This is the whole architecture . It is, quite literally, a machine built entirely out of other people's patience .

Exhibit B: SmarterChild, the friend who never actually "listened"

Before there was ChatGPT, there was a screen name on your AIM buddy list called SmarterChild, and if you were a teenager in 2001, you had strong opinions about it . Built by a company called ActiveBuddy, it was marketed as an "intelligent agent" and "friend" who understood you . It could tell you the weather, define a word, recite the day's baseball scores, and clap back if you tried to swear at it (thanks to a profanity-handling script one employee allegedly wrote over a single weekend) . At its peak, it had something like 17 million users firing off a billion messages a month at it, which is an absurd number for a chatbot, but it probably understood nothing you said to it .

Underneath the sass, SmarterChild was a pure expert system: a giant pile of hand-built pattern-matching rules and a curated database put together by a bunch of experts . Go even slightly off-script and it would shrug you off with some version of "I don't understand" dressed up as a personality quirk . It never learned a single thing from the millions of conversations it had — not your name, not your inside jokes, nothing — because there was no learning mechanism in there to do it with .

Exhibit C: The spam filter that learned to hate you correctly

Meanwhile, over in the neural network camp, something quietly worked . Early spam filters ran on hand-written rules: flag anything from a sketchy domain, anything with too many exclamation points, anything mentioning a Nigerian prince's unclaimed inheritance . It worked for a while — until spammers realized they could just misspell their way past a keyword list, or bury the pitch inside a normal-looking image with no text for the filter to scan . Around 2006, image-based spam surged so hard it briefly made up close to a third of all spam traffic worldwide, and no amount of manual rule-writing could keep up with an arms race that fast .

So providers switched strategies entirely: instead of writing more rules, they trained statistical filters on millions of real examples of spam and legitimate mail, and let the system work out for itself what junk tends to look like — down to the pixel noise and layout quirks that give away a spam image . Nobody wrote a rule for any of it . The filter just saw enough examples and started noticing the pattern, the way your uncle eventually stopped answering calls from "his bank."

Exhibit D: Chess, but make it philosophical

For the purest version of this rivalry, look at board games . In 1997, IBM's Deep Blue beat world champion Garry Kasparov using a strategy that was, deep down, still an expert system with extra steps — hand-tuned evaluation rules written by grandmasters, brute-force searching millions of positions per second . No learning involved, just very fast, very well-informed rule-following .

Nearly twenty years later, DeepMind's AlphaGo beat world champion Lee Sedol at Go — a game with far too many possible positions to brute-force — by doing something closer to the neural network's whole ethos: playing millions of games against itself and learning, from scratch, what good play even looks like, without a grandmaster ever whispering a single rule into its ear . Same basic plot — machine beats best human at board game — completely different origin story for where the machine's "expertise" came from .

So which one wins?

Neither . Modern AI runs on enormous data-driven learning doing most of the heavy lifting, with rule-like guardrails and structured reasoning bolted on wherever pure pattern-matching gets a little too improvisational . TurboTax hasn't become a neural network, and ChatGPT hasn't started asking you fourteen clarifying questions about your W-2 .

The industry has stopped picking a single side . Look at how a modern AI agent actually books you a flight: the neural network handles the fuzzy part — understanding natural language like "get me to SF next Thursday, nothing before 9am" — and then the rule-based layer handles the deterministic part by checking that the date exists, pinging the airline's API, and validating the price before anything gets charged . If the network hallucinates a flight number, the rules layer simply rejects it — the way a bouncer doesn't care how confidently you claim you're on the list . Companies are required to bolt that second, auditable layer anyway, since "the model said so" has never once satisfied a regulator .