You've got a handful of tasks you're considering handing to AI, and no way to tell which ones are safe to try this week and which still need a person. AI just solved 90 of the top 500 hardest open problems in mathematics, so the line keeps moving. There's a way to check any task yourself, without guessing.
In this guide: a four question test that sorts any task by how much disorder it carries, and tells you what to do with it.
Start with the basics
The idea behind this comes from physics, borrowed by a distinguished scientist at Google: automation speed runs in inverse proportion to how much entropy, disorder, a field has. Math and code are tidy and rule-bound, so AI already lives there. The messier, noisier and more human a field gets, the later AI arrives. That's not a guess, it's an order.
The disorder check for your task
Why this works: the more disorder in a domain, the slower AI moves through it. Math and code come first in line. Everything messier comes after, in that order.
- Does it have one answer you can check is correct?Code that runs or doesn't, a calculation with a clean solution. If yes, AI is already there.
- Are the rules fixed, but the real-world measurement is messy?Lab analysis, sensor readings, anything with clean physics but noisy data. AI is good for a first draft.
- Do conditions keep shifting, and does the outcome depend on what people think?Pricing, sales forecasts, markets. AI helps you read the data, it doesn't decide for you.
- Are the stakes high, the information incomplete, and the judgment calls never written down?Diagnosis, contracts, policy calls. This stays human for now.
Try it on two real office tasks. Matching supplier invoices to purchase orders: there's one answer you can check, it matches or it doesn't, that's question 1, and AI can already run it alone. Deciding whether a customer keeps their credit line when their payment history is mixed: there's no formula for that, that's question 4, and it stays with a person.
This isn't theoretical. Last week OpenAI published 372 new mathematical results and fully solved 90 of the top 500 open problems in the field. Alex Kontorovich, a math professor at Rutgers, reacted: “If a human did this, it would be an instant Fields Medal, no questions asked.” That's the signal that the first zone on this list, the one with a single checkable answer, has already crossed the line.
Next ten minutes: write down five tasks you've been considering for AI, run each one through the test, and start with whichever one answered yes on question 1.