There’s a line of thinking you hear everywhere right now: AI can do everything, so just hand it the whole job. Give it your emails, your code, your spreadsheet, the whole mess, and let it sort things out. It sounds obvious. Why keep doing things the hard, manual way when there’s a machine that seems to understand anything you throw at it?

I want to gently take that idea apart, because underneath it sits one small concept that almost nobody explains in plain language. Once you have it, a lot of confusing things about AI suddenly make sense: why it gives a different answer when you ask twice, why people say “don’t trust it with money or medicine,” and why engineers keep building boring old-fashioned code around the clever AI instead of just letting the AI run free.

The concept has two intimidating words attached to it: deterministic and probabilistic. Ignore the words for a minute. Here’s the whole idea in one picture.

A calculator and a very smart friend

Punch 847 × 39 into a calculator. You get 33,033. Do it again tomorrow, on a different calculator, on the other side of the world: 33,033. A hundred years from now, same answer. The calculator does not have moods. It does not get tired. It does not “think it’s probably around 33,000.” It follows fixed rules and lands on the exact same spot, every single time.

Now ask your smartest friend to do 847 × 39 in their head. They’re sharp, so they’ll probably get it right, or very close. But they might say “about 33,000.” They might slip and say 32,900. Ask them again next week and they’ll phrase it differently, maybe get it exactly right this time. They’re brilliant and useful and they can also do a thousand things the calculator can’t. They just don’t guarantee the same exact answer every time.

The calculator
847 × 39 = ?
Mon33,033
Tue33,033
Next yr33,033
Same input, same answer, forever. This is deterministic: fixed rules, no surprises. Boring. Trustworthy.
The smart friend
847 × 39 = ?
Mon"about 33,000"
Tue"33,033, I think"
Next yr"~32,900?"
Usually great, occasionally off, never phrased the same. This is probabilistic: it works in likelihoods, not guarantees.
That's the entire idea. Deterministic = same answer every time. Probabilistic = a very good guess that can vary. Both are useful. You just want the right one for the job.

That’s it. That’s the whole concept. Deterministic means “follows fixed rules and gives the identical result every time.” Probabilistic means “works out what’s most likely to be right, which is usually excellent and occasionally wrong, and can come out a little different each time.”

Here’s the same thing as a picture. Ask each one the same question five times in a row and watch what the answers do.

Answer to 847 × 39, asked 5 times (correct answer: 33,033)
33,033
33,033
33,033
33,033
33,033
33,033
~33,000
32,900
33,033
33,100
1234512345
Calculator (deterministic) Smart friend (probabilistic)
Left five bars: flat, identical, boring, exactly right. Right five: mostly great, wobbling around the true answer, never quite the same twice. That wobble is the whole story. It's not the friend being dumb; it's what a guessing machine does.

Today’s AI, the kind behind chatbots and coding assistants, is the smart friend. It is astonishingly capable. It is also, at its core, a machine for predicting the most likely next thing to say. That prediction is usually right. But “usually right, and I can’t promise the same answer twice” is a very different tool from “exactly right, every time.” Neither is better. They’re for different jobs.

Why you already trust this instinct in real life

You don’t need to be told which tool to use. You already know it, you’ve just never named it.

  • You’d let a friend recommend a restaurant. You would not let them do the restaurant’s payroll by memory. Recommendations are a “good guess” job. Payroll is an “exact, every time” job.
  • You’d ask a friend to proofread your speech for tone. You’d use a spellchecker’s dictionary to confirm a word exists. Judgment vs. a fixed rule.
  • You’d trust a chef cooking by taste to make your dinner delicious. You’d want a pharmacist measuring your medicine to follow the exact recipe, no improvising.

The dinner can be a little different each night and that’s part of the joy. The medicine cannot. The stakes and the need for sameness decide which kind of machine you want. Hold that thought, because it’s the key to the whole AI question.

The same split, now in technology

Here’s the part that surprises people: your phone and laptop are full of both kinds, quietly working together. You never noticed because it just works.

Your calculator app, and the "search" box
The calculator is pure deterministic: rules in, exact number out. But the moment you type a half-remembered question into a search bar and it *guesses what you meant*, corrects your typo, and ranks a billion pages by "what you probably want", that's the probabilistic side stepping in. Same device, two different kinds of thinking.
Maps: the map vs. the arrival time
The road map itself is fixed data, this street connects to that one, deterministic. But "you'll arrive at 6:42pm" is a prediction built from traffic patterns and probabilities. That's why the map is never wrong about where a road is, but the arrival time keeps shifting. Two different machines under one app.
Translation apps
Older translators used fixed dictionaries and grammar rules, deterministic, and produced stiff, robotic sentences. Modern AI translators guess the most natural phrasing, which reads far better and is occasionally, confidently wrong in a funny or embarrassing way. The upgrade from "rigid but literal" to "fluent but fallible" is exactly this deterministic-to-probabilistic shift.
Autocorrect and predictive text
Every time your phone finishes your sentence or "fixes" a word you meant to keep, you're watching a probabilistic machine guess the likely next word. Brilliant most of the time. The source of a thousand embarrassing texts the rest of the time. That's not a bug in the usual sense, it's what a guessing machine is.
You've been living with both kinds your whole life. The new part isn't the existence of guessing machines. It's that the guessing ones suddenly got so good that people started asking: why not use them for everything?

Now, the big question: “so why not just give everything to AI?”

This is the honest heart of it. If the smart friend is so good now, better than most experts at most things, why keep the boring calculator-style machines around at all? Just hand the AI the whole job.

The answer is not “AI is bad.” AI is genuinely amazing. The answer is: some jobs need the same right answer every time, and a guessing machine, no matter how smart, can’t promise that. Let me make it concrete with jobs people actually do.

AI helping you write code. This is one of the most popular uses today, tools like GitHub Copilot, Cursor, and Claude Code. You describe what you want, the AI writes the code. It’s genuinely wonderful and saves hours. But it’s the smart friend: it writes code that looks right and is usually right, and every so often it invents a function that doesn’t exist, or gets an edge case subtly wrong, with total confidence. That’s why every serious team still runs the AI’s code through boring, deterministic tools: a compiler that checks the code is even valid, automated tests that pass or fail the same way every time, a linter with fixed rules. The AI is the creative friend brainstorming; the tests are the calculator checking the math. You want both. Handing the whole thing to the friend with no checker is how you ship a confident mistake.

AI reviewing code. A newer, related use: let the AI read a proposed change and comment on it, like a helpful colleague looking over your shoulder. It catches real things and it’s a great extra set of eyes. But you would not let it be the only gate, because it’s still guessing. It might wave through a genuine bug or flag something harmless as dangerous, and not the same way twice. So the AI review sits alongside the deterministic checks (does it compile, do the tests pass, does it break a rule we’ve written down), not instead of them. Friend gives an opinion; the fixed checks give a verdict.

Moving your money. Ask an AI to help you understand your spending, great, that’s a judgment-and-explanation job. But the actual transfer of $4,000 from one account to another must be deterministic to the penny, logged, repeatable, reversible. Nobody sane wants “the AI moved roughly the right amount.” The bank’s core runs on old-fashioned, exact, boring machinery for exactly this reason.

Medicine and dosages. An AI can be a superb assistant, spotting patterns in a scan, suggesting things a tired human missed. But the dose calculation, the “give 5mg, not 50,” is a place you want fixed rules and a human double-check, not a confident guess that varies.

See the pattern? In every case the winning move is not “AI or not AI.” It’s AI for the parts that need judgment and language, and boring deterministic machinery for the parts that must be exact and repeatable. “Just give everything to the AI” fails not because AI is weak, but because it quietly hands the exact-every-time jobs to a machine whose whole nature is to guess.

If you want it all in one place, here are the two side by side.

QuestionDeterministic (the calculator)Probabilistic (the smart friend / AI)
Same answer every time?Yes, alwaysNo, it can vary
How it decidesFollows fixed rulesGuesses what's most likely
When it's wrongSame way every time (a findable bug)A different way each time (hard to pin down)
Great atExactness, repeating, promisesLanguage, judgment, handling messy input
Bad atAnything fuzzy or open-endedBeing exact and identical every time
Everyday exampleCalculator, bank transfer, spellcheck dictionaryChatbot, translation, "you'll arrive at 6:42"
Trust it alone with money?YesNo, check it first
Neither column is the "good" one. They're different tools. The mistake is using the right-hand one for a left-hand job.

The secret most people miss: it’s a dial, not a switch

Here’s the part that makes you sound wise at a dinner party. Almost nothing real is purely one or the other. Most good systems are a mix, and the skill is knowing which parts to make exact and which to let the AI guess at.

Exact every time (deterministic)(probabilistic) Smart guess
exactA calculator, a bank transfer, a password check. Must be identical every time. Zero room for "probably."
exactRunning code and its tests. Pass or fail, the same way each run, or you can't trust anything.
mixA search engine. Fixed data underneath, a smart guess about what you meant on top.
mixAn AI coding assistant. Creative guessing, checked by exact tools before anyone trusts it.
guessWriting an email, brainstorming names, translating a poem. You want variety and judgment. Sameness would be worse.
guessRecommending a movie or a restaurant. "Here's a good guess for you" is exactly the point.
Not two boxes, a slider. Good design is putting each part of a job at the right spot on this line, not shoving everything to one end.

And notice the bottom of that list: for some jobs, the guessing is the whole point. You don’t want an email-writer that produces the identical sentence every time, or a movie recommender that suggests the same film to everyone. There, variety and judgment are features, not flaws. Probabilistic isn’t the “worse” option. It’s the right option when you want creativity, nuance, and language. It’s the wrong option when you need a promise.

The way the pros actually build it

So what do good engineers do? They don’t pick a side. They combine the two on purpose: let the exact machinery do the parts that must be reliable, and let the AI do the part it’s uniquely good at, judgment and language, on top.

exactGather the real facts with fixed, reliable code. Pull the exact numbers, the real records, the actual code, no guessing at this stage.
ai / guessLet the AI read those facts and explain them in plain language. This is what it's brilliant at: turning a pile of facts into something a human understands.
exactCheck the AI's work against the facts before trusting it. Did it stick to what's real? Fixed rules and tests catch it if it wandered off.
humanA person makes the final call on anything that matters. The AI advises; it doesn't get to press the irreversible button alone.
The recipe behind most trustworthy AI products: exact machinery holds the truth, the AI makes it human, exact machinery checks the AI, a person decides. Nobody hands the whole job to the guesser.

This is why, when you use a well-built AI product, it often shows you where its answer came from, a real source, a real number, a real line of code. That’s the exact-machinery part quietly keeping the smart friend honest.

The one thing to remember

If you take a single sentence from this: when an AI gives you a different answer the second time you ask, that’s not it malfunctioning, that’s what it fundamentally is. It’s a guessing machine, an extraordinary one, and guessing machines vary. That’s a strength for writing, brainstorming, translating, and explaining. It’s a risk for anything that has to be exact and the same every time, like money, measurements, and things you can’t undo.

There’s one last picture worth burning into memory: the two kinds are wrong in different ways. Run each one 20 times and watch how the mistakes behave.

Deterministic: 20 runs, same job
Probabilistic: 20 runs, same job
The deterministic row is all one colour, right or wrong, it's the same every run, so a mistake is a bug you can find once and fix forever. The probabilistic row is right most of the time, then slips somewhere different each go (the off-colour squares). That's why it's brilliant and why you can't fully trust it alone: you can't predict which run will be the one that's off.
"Wrong the same way every time" vs "wrong a different way each time." That single difference decides where each kind belongs.

So the smart question is never “AI or not AI.” It’s the same question you already ask in real life without thinking: does this job need the calculator, or the clever friend? Most big jobs need both, arranged so each does the part it’s actually good at.

The “just give everything to the AI” crowd isn’t wrong to be excited. They’ve just skipped the small, old idea that quietly runs underneath all of it. Now you have it, and I promise you’ll start noticing it everywhere.

Want to go a little deeper?

All explained in my own words above, but if you want to explore these ideas further, here are solid, freely available starting points:

  • What “probability” even means, gently. Khan Academy’s intro probability lessons are free and genuinely beginner-friendly: khanacademy.org.
  • How AI actually “guesses” the next word. 3Blue1Brown’s visual explainer series on neural networks and language models is the clearest free video introduction I know: youtube.com/@3blue1brown.
  • Why AI sometimes states wrong things with confidence. Search for “AI hallucination” explainers; most major AI labs publish plain-language write-ups on why it happens and what reduces it.
  • The deterministic side, in one word: algorithms. If “fixed rules, same answer every time” grabbed you, that’s the world of algorithms. Any beginner “intro to how computers follow instructions” resource is a good next step.

If a young person or a not-so-technical person in your life is confused about all the AI noise, this is the one idea I’d hand them first. Everything else gets easier once the calculator and the clever friend are clear in your head.

← Back to blog