“AI mainly boosts people who already understand what they are doing.” That claim is substantially right, but it misses something important. A person who “gets it” is not necessarily someone who entered the room carrying a warehouse of expert knowledge in their head.
What matters more is whether they can look at an output and judge: this is useful, this is suspicious, the comparison set is too small, or this is not actually done. And those standards do not all have to exist in advance. If you do not know the standard, you can ask, “What do other companies do?”, “What is the industry norm?”, “Where have similar attempts failed?”, and “What is the strongest counterargument?” In other words, you can research the evaluation criteria themselves.
At that point, AI starts to look less like a knowledge vending machine and more like a sparring gym for comparing, probing, falsifying, and increasing the resolution of thought.
1. “People who already understand get stronger” is only half the story
It helps to think of AI as a multiplier. If the goal is fuzzy, it can accelerate fuzziness. If the user has a sharp question and useful criteria, it can accelerate exploration. That is why experienced practitioners often appear to get disproportionate value from strong AI systems.
But prior knowledge is only one thing being multiplied. The ability to find comparisons, notice anomalies, challenge a hypothesis, and define what “done” means can also be amplified. As AI takes on more of the searching and organizing, the human ability to decide what evidence would actually distinguish right from wrong becomes more valuable.
A deliberately rough formula would be:
Practical AI multiplier ≈ knowledge × evaluation criteria × comparison skill × falsification skill × number of iterations
There are domains where near-zero knowledge is genuinely limiting. But having enormous knowledge while accepting every polished answer with “sounds right” is not a particularly strong multiplier either.
2. “Understanding” is less about storing facts and more about judging outputs
The idea becomes clearer if we split “understanding” into five abilities.
- State the objective: What problem are we solving, and what final state do we want?
- Judge quality: What counts as passing, and what counts as failure?
- Find the bottleneck: Where is the system actually constrained, rather than merely showing symptoms?
- Detect suspicious claims: Do you have enough of a mental model to stop and ask, “Is that really true?”
- Define done: When should the research, debugging, or revision stop?
You do not need expert-level mastery of all five. The critical shift is from “this text sounds plausible” to “this output satisfies or fails the objective for these reasons.”
That difference is easy to underestimate. Knowing the answer and being able to evaluate an answer are different capabilities. AI makes the second one increasingly important.
3. You can import evaluation criteria: “What do others do?” is an unusually powerful question
Here is the key twist: if you do not already have a standard, research the standard.
When entering an unfamiliar business process, you do not have to stop at “I do not know the best practice.” You can ask AI to gather and structure questions such as:
- What do other companies do?
- Is there an industry standard?
- How do top performers differ from ordinary firms?
- Where did failed implementations break?
- How do adjacent industries solve the same constraint?
- What is the minimum required by regulation or formal standards?
- What do users consistently complain about?
Look at one company and almost anything can seem normal. Look at five and differences appear. Add failures and you start seeing why the differences matter.
What do others do? is therefore not just a request for trivia. It is a mechanism for generating a grading rubric you did not previously possess.
4. AI is not just a knowledge highway; it is a machine for acquiring comparison axes
In an unfamiliar domain, AI can shorten the distance from “complete beginner” to “I at least know what must be checked.” A useful sequence is:
unknown → overview → comparison → hypothesis → examples → counterarguments → exceptions → primary evidence → evaluation criteria → provisional judgment
The goal is not instant expertise. It is to reach the point where you can ask better questions, notice weak evidence, and compare alternatives intelligently. Once you get there, the next round of research improves dramatically.
All of this was possible before AI. It was simply expensive. Finding ten competitor cases, learning the terminology, reading failures, and crossing into adjacent industries could take substantial time. AI reduces the friction of searching, reorganizing, and translating across contexts. So the winner is not only “the person who already knows.” It can also be the person who can move into a state of knowing much faster.
5. The core skill is probing: do not go home after answer number one
A major difference in AI use appears when the first answer is treated as the start rather than the finish. Instead of saying “interesting” and leaving, keep pressing.
“What is the evidence?” “What changes concretely?” “What do other firms do?” “Is this actually common?” “What are the exceptions?” “Show the opposite case.” “Are you mixing correlation with causation?” “What is the real bottleneck?” “What would we observe if this hypothesis were false?”
A generic first answer becomes relative once comparison is added. Exceptions reveal boundary conditions. Falsification exposes weak points. Primary evidence shows which parts survive contact with reality. Only then does a plausible answer start becoming decision material.
People who use AI well do not necessarily write a masterpiece as their first prompt. They increase resolution mercilessly on turn two, turn three, and turn ten. That is the core behavior.
6. Why the same level of probing is difficult with humans
Repeated questioning between people creates costs beyond information retrieval. It consumes someone’s time, may feel like distrust, can touch status or pride, and produces explanation fatigue. Those are interpersonal frictions, not necessarily intellectual objections.
A good colleague or expert may happily engage for quite a while. Still, if you fire “evidence?”, “exceptions?”, and “isn’t that assumption wrong?” thirty times in a row, the atmosphere will eventually change even if every question is legitimate.
In intentionally exaggerated form:
- Question 5: one eyebrow moves.
- Question 12: the sigh becomes shared meeting-room infrastructure.
- Question 25: “Do we really need this right now?” enters the chat.
- Question 30: ending the meeting becomes a more urgent KPI than answering the question.
That is not because humans are defective. Humans have time, emotions, roles, and other responsibilities. Nobody exists solely to answer an infinite recursive query loop.
7. AI can become an all-you-can-probe sparring gym
With AI, much of that interpersonal friction is dramatically lower. You can return on question twenty-eight with “Okay, but what is the evidence?” You can jump back fifteen minutes and ask, “Doesn’t your previous explanation contradict this?” without creating a new office legend about yourself.
This means the value of AI is not merely the amount of information it contains. It also lowers the cost of repeated thinking. You can attack the same hypothesis from another angle, add competitors, reverse assumptions, rewrite the explanation for a novice, then rebuild it for an expert.
But “no social cost” should not be interpreted literally as zero cost. Iterations still consume time. You can spend thirty rounds moving confidently in the wrong direction. And fluent AI output can create its own cognitive pull. This is not a free, infinite truth machine.
8. A stronger boost can also amplify a wrong idea
A multiplier has an uncomfortable property: it can multiply error too. If your starting hypothesis is wrong and you ask only for supporting evidence, AI can help turn a bad idea into a beautifully organized bad idea at impressive speed.
Four failure modes are especially useful to watch:
- Automated confirmation bias: collecting only cases that support the desired conclusion.
- Fluent wrongness: assuming strong prose implies strong evidence.
- Benchmark cargo culting: copying what another company does while ignoring different constraints.
- False consensus: seeing a handful of examples and declaring, “This is what everyone does.”
The countermeasure is simple and inconvenient, which is exactly why it works. Ask, “What is the strongest argument against this?” “Find cases that go the other way.” “What does the primary source actually say?” “Under what conditions does this comparison fail?” Use AI to attack your own conclusion, not only to decorate it.
A booster without steering just reaches the wall faster.
9. A strong loop is comparison → falsification → primary evidence → re-judgment
A practical workflow looks like this:
- Place a hypothesis: What currently seems to be the cause?
- Build a comparison set: Other companies, countries, systems, industries, and historical cases.
- Extract evaluation criteria: Which differences actually appear to matter?
- Make a provisional judgment: Choose the current leading explanation.
- Try to falsify it: Search for contrary evidence, exceptions, and alternative hypotheses.
- Move toward primary evidence: Standards, original documents, measurements, official records.
- Update the judgment: Do not protect the first hypothesis out of pride.
- Define done: Stop when additional research is unlikely to change the decision.
A compact question set makes this reusable:
| Need | Ask AI |
|---|---|
| Comparison | “What do other companies, countries, or methods do?” |
| Normality | “Is this common or unusual? What is the relevant population?” |
| Exceptions | “Under what conditions does the opposite happen?” |
| Cause | “What is the bottleneck rather than the visible symptom?” |
| Falsification | “What evidence would disprove this hypothesis?” |
| Completion | “What would be enough to verify before deciding?” |
This is less “ask AI for the correct answer” and more use AI to train your own judgment system.
10. The new “person who gets it” is not the person who knows everything
A more useful definition for the AI era is:
Someone who recognizes what they do not know, gathers the right comparisons, builds or imports evaluation criteria, tries to falsify hypotheses, judges outputs independently, and updates quickly when the evidence changes.
That is not the same as carrying an encyclopedia in your head. It is closer to knowing where your ignorance is and what evidence would make the problem judgeable.
This is why unfamiliar domains become more accessible. Even if experts start with a large knowledge advantage, a fast comparison-and-falsification loop can move a beginner out of the “I do not even know what the real questions are” state surprisingly quickly.
None of this makes experts obsolete. Deep tacit knowledge, field experience, and accountable judgment remain valuable. What changes is how quickly a non-expert can build the scaffolding required to understand an expert, challenge weak claims, and know what to investigate next.
11. Conclusion: do not merely ask AI for answers; keep increasing resolution
“AI makes people who already understand even stronger” is true. A more precise version is better:
AI strengthens not only people who already possess evaluation criteria, but also people who can build them quickly, import them from outside, and repeatedly pressure-test their conclusions.
You do not need to know everything at the start. Keep asking: “What do others do?” “Is this normal?” “What are the exceptions?” “What is the evidence?” Then go one step further: “What would make me abandon this hypothesis?” That is how judgment grows even in a new domain.
Do this thirty times to a human and the oxygen in the meeting room may feel thinner. AI will still entertain question twenty-nine. The important part, however, is not that you are allowed to ask twenty-nine times. It is whether, when your hypothesis breaks on turn twenty-nine, you are willing to rebuild it on turn thirty.
The point is: pressure-test the output. Do not merely receive the answer. Push it until the weak parts break, then use what remains as decision material. That is a large part of what the “boost” really is.
