Verifying AI Output

You give an AI assistant your monthly budget notes and ask for a summary. It replies with a tidy breakdown, a total, and a few suggestions. You paste the total into the spreadsheet you share with your partner. It is 10 too low. Every amount in the reply was copied correctly. The addition was wrong.

Nothing in the reply looked off. The sentence with the wrong total sounded exactly as confident as the sentences that were right. That is the reason to verify: you cannot tell from the writing which parts to trust.

#Checking means comparing with something else

Reading a reply again and nodding is not a check. Neither is asking the model "are you sure?", for the reasons in Why Models Hallucinate. A real check compares the reply with something outside it: your original data, a calculation you run, an official document, a test that passes or fails.

The routine has two steps.

1. Split the reply into claims. A claim is any sentence that could be true or false: a number, a name, a date, "this function handles empty input", "the report says X".
2. Check each claim the way that suits its type. Numbers get recomputed. Facts get looked up in a primary source. Code gets run. Citations get opened and read.

#Watch a reply get checked

Here is the budget request and the reply. Read the reply once and see if anything jumps out.

Nothing jumps out, and that is the point. Step through checking its six claims one at a time.

#Numbers: recompute, never re-read

Language models produce numbers the same way they produce words, one token at a time. Simple arithmetic often comes out right, and sometimes it does not. When a number matters, recompute it from the inputs. For more than a couple of values, a few lines of code beat a calculator because you can rerun them.

const spending = { rent: 950, groceries: 312, transport: 86, phone: 45 };
const pay = 2600;
 
const total = Object.values(spending).reduce((sum, x) => sum + x, 0);
console.log("total:", total); // 1393
for (const [name, amount] of Object.entries(spending)) {
  console.log(name, ((100 * amount) / total).toFixed(1) + "%");
}
// rent 68.2%, groceries 22.4%, transport 6.2%, phone 3.2%
console.log("left over:", pay - total); // 1207

spending = {"rent": 950, "groceries": 312, "transport": 86, "phone": 45}
pay = 2600
 
total = sum(spending.values())
print("total:", total)  # 1393
for name, amount in spending.items():
    print(name, f"{100 * amount / total:.1f}%")
# rent 68.2%, groceries 22.4%, transport 6.2%, phone 3.2%
print("left over:", pay - total)  # 1207

If you ask the assistant to do the maths, ask it to write code like this and run it in a tool, or to show each step so you can check the steps. A number with no working is the hardest kind to trust.

#Facts, code and citations

#Facts

Names, dates, version numbers, laws, prices and quotes need a primary source: the official documentation, the original report, the law itself. A page that repeats the claim is weaker: it may copy the same mistake. For software, check the docs for the version you actually use. A function that existed three versions ago, or only in a different library, is a common wrong answer.

#Code

Run it. Then try the inputs the example did not cover: empty, one item, duplicates, negative numbers, very large values. Coding With AI shows how to turn those cases into tests before you accept the code.

#Citations

A citation is three claims at once: the source exists, it says what the reply says it says, and it is about the same thing. Check all three.

• Open it. A link that fails to load, or a paper you cannot find by its title, is a red flag.
• Find the sentence. Search the page for the specific number or phrase. A source on the right topic that does not contain the claim does not support it.
• Read around it. A quote can be real and still mean the opposite in context, or describe a different year, country or product.

#How much to check

You do not need to verify a brainstorm of party themes. You do need to verify a figure going to your manager, a dose, a legal deadline or code going to production. Before checking, ask: if this is wrong, who notices, and what does it cost? Then check the claims where the answer is "someone else, and a lot" first.

#Try it yourself

#Common mistakes

#Check yourself

#What to practise next

Take the last long AI reply you used for something real and split it into claims. Check the two that would hurt most if wrong. Then continue with Writing a Good Brief, which shows how to ask for work that is easier to check in the first place.