Practical AI · 3-minute read
There is a subscription for almost everything now. Fortunately, asking “How do we know that?” has not yet moved behind a premium tier.
Critical thinking does not mean rejecting every AI answer or turning a group chat into a courtroom. It means giving a claim the amount of scrutiny its consequences deserve.
A dinner-party theme and a major purchasing decision do not need the same review process. Neither benefits from treating confident wording as proof.
The question before the spreadsheet
Sundial Grove Packing, a fictional agricultural packing business, is comparing two packaging suppliers. A manager asks AI which option is more sustainable. The answer favors one supplier and sounds reassuringly complete.
But “more sustainable” has not been defined. Does the business mean less material, less product damage, a locally available disposal route, or lower transport emissions? What supporting information was supplied? What is merely a vendor’s claim?
Those questions may change the comparison before anyone changes the supplier.
AI can help organize the evaluation once the team agrees on criteria. It can list unanswered questions, build a comparison table and point out where a conclusion depends on an assumption. The team still needs to obtain and examine the relevant evidence.
Four questions worth keeping
What exactly is the claim? Replace broad labels such as “better,” “safe” or “efficient” with something specific enough to examine.
What supports it? Find the original source and check whether it actually supports this claim in this setting.
What are we assuming? Look for missing costs, unusual conditions and exceptions that a tidy answer might overlook.
What would change our minds? Name the evidence that would weaken or reverse the conclusion. Otherwise, research can become shopping for agreement.
NIST’s AI Risk Management Framework emphasizes interpreting AI outputs in context and evaluating performance. The questions above are my everyday version of that habit, not a certification process.
Try this prompt
Help me evaluate this claim: [claim]. The decision is [decision], and our criteria are [criteria]. Using only the material provided, list supporting evidence, unsupported assumptions, missing information and plausible alternative explanations. Explain what would change the conclusion. Do not invent sources or treat a supplier’s statement as independently verified.
Check the resulting table against the actual documents. Use a tool approved for the material, and involve the person who understands the purchase. If the evidence is incomplete, a useful result may be a better question to send the supplier.
You can use AI to make your thinking more organized. You do not need to let it decide when you are finished thinking.
That feature is still included in the human plan.
Your turn: What question do you wish someone had asked before the last “obvious” decision at work?
Background: NIST AI RMF 1.0, MEASURE. The company, purchasing scenario, four questions and prompt are original editorial examples.
