How to check AI's work without redoing it yourself

AI can state a wrong figure, a made-up source or a broken link with complete confidence. That does not mean you have to redo its work to be safe. It means you check it the way you would check a capable new colleague: ask for evidence, check harder when more is at stake, and get a second pair of eyes on anything that really matters.

The trap: checking by redoing

When people first find a mistake in AI work, they often swing too far. They start redoing every task by hand “just to be sure”. At that point the AI is saving them nothing.

The goal is different. Your job moves from doing the work to reviewing it. A good review is quick, targeted and in proportion. Here is how.

Ask for proof, not promises

“I’ve checked it and it’s correct” is a promise. You want proof.

• Ask how it knows. A plain “How do you know this is right? Show me.” often surfaces a guess dressed up as a fact.
• Ask for sources. For facts, ask for links. Then open one or two yourself. You do not need to open all of them. A quick spot check tells you a lot.
• Ask it to test. For anything that can be tested, have the agent test it. Add the numbers up another way. Open the file it made. Try the link. Compare the totals with the original.

Try saying:

Before you hand this over, check it: add up the totals a second way, confirm every link opens, and tell me anything you could not verify.

That last part matters most. “Tell me anything you could not verify” gives the agent permission to say what it is unsure of, instead of smoothing over it.

Check in proportion to the stakes

Not every piece of work deserves the same scrutiny. Checking a lunch order like a contract wastes your time. Checking a contract like a lunch order is how mistakes reach clients.

A simple scale works well:

• Low stakes: an internal draft, a brainstorm, a rough summary for yourself. Read it.
• Medium stakes: an email to a client, a summary for your manager. Read it, and ask for sources on any facts.
• High stakes: a figure in a contract, advice someone will act on, a payment. Ask for proof and sources, a test where possible, and a review by another model family.

Decide which level a job is before you start. Then you know how much checking to plan for, and you do not over-check the easy work or under-check the hard work.

Get a second opinion from another family

Models from the same family tend to share blind spots, because they were trained in similar ways. Ask a model to check its own work and it often agrees with itself.

A model from another family is far more likely to spot what the first one missed. Anthropic’s Claude models and OpenAI’s GPT models are trained very differently, so they notice different things. If Claude did the work, have GPT check it, and the other way round.

Try saying:

Have a GPT agent review this contract summary on its own and list anything wrong or missing. Then tell me where you agree and disagree.

Two habits make second opinions much more useful.

Keep it independent

The reviewer should not see how the first model got there. If it reads the first model’s reasoning, it tends to follow it, and you lose the point of asking twice.

Assume there is a mistake

Reviewers find more when you tell them something is wrong. It turns “looks fine to me” into an actual search.

Try saying:

Assume there is at least one mistake in this. Have a helper from another family, such as GPT or Gemini, find it, without seeing how you got here.

For work that must be right, use a council

For the highest-stakes work, one review is not always enough. A council takes the second opinion further: reviewers from different model families score the work, the agent fixes what they found, and they review it again, round after round, until both are satisfied.

This works well for things you will send to many people or rely on for a long time: a proposal template, a policy, a quote that new customers must understand without calling you. The key is to say clearly what “better” means, so the reviewers score against what you actually care about.

Set limits too. The first rounds usually gain a lot, and later ones less and less. A sensible limit on rounds and time stops a review going on long after it has stopped helping.

Watch for extras, not just errors

Mistakes are not the only thing to check. Some models, Claude in particular, fill gaps and often do more than you asked. That is usually helpful. But if you asked for a spelling check and the pricing section changed too, that is a problem.

Try saying:

When you are done, list every change you made.

A list of changes lets you spot any extras at a glance, without reading the whole document side by side.

When something keeps coming out wrong

If the same mistake keeps coming back, the cause is often one wrong belief early on. The agent assumed something, built on it, and every fix has been built on the same mistake.

Try saying:

Stop for a moment. List everything you are assuming about this problem: about the data, the files, what I want, and how it should work. Mark the ones you have not checked.

One of the unchecked ones is usually the culprit. Also tell it exactly what you can see. “It’s still wrong” helps nobody. “The March total should be $48,210. Yours is $52,900. Here is the bank statement for March.” gets it fixed.

You own the result

Whatever the agent produces, the person who sends it on is responsible for it. Read anything before it goes to a client, a colleague or into a decision. The agent is fast. You are the judgment. Keep that division and you get the speed without the risk.

A quick checking routine

1. Decide the stakes: low, medium or high.
2. Ask the agent to check its own work first, and to say what it could not verify.
3. For facts, open one or two sources yourself.
4. For anything high stakes, get an independent review from another model family.
5. Ask for a list of changes so you can spot extras.
6. Read it before it leaves your hands.

How Interweave OS does it

In Interweave OS you can ask your chat’s agent to have a subagent from another model family review any piece of work, independently, without leaving the chat. The Second opinion button under any agent reply asks one reviewer from another family whether it agrees, what is wrong, worst first, and what it would change, without changing anything. For work that must be right, Improve with council has reviewers from different families score the work out of 10 in each category, round after round, until both reach your pass mark or the round and time limits you set. Everyone in the chat can watch the scores, and anyone can stop it, keeping the best version.