Trust But Verify: Catching AI Mistakes Before They Cost You

Everything in this path makes you faster. This lesson makes sure faster does not mean wrong. AI writes smooth, confident text whether or not it is true. The more you rely on it, the more a rare mistake costs you. The goal is not to distrust AI. The goal is to know where it fails, check in proportion to the risk, and keep private data private.

Why AI gets things wrong

• Made-up details: it fills gaps with believable facts, quotes, references, statistics or sources that do not exist.
• Old or missing knowledge: it may not know recent events, or your company's private facts.
• Misreading your request: a vague prompt gets a confident answer to a different question.
• Slips with numbers and counting, especially in long tables and multi-step sums.
• Agreeing with you: models tend to tell people what they seem to want to hear, even when the starting idea is flawed.
• Hidden instructions: content it reads can steer it (see the lesson on browser agents).

A checklist for checking

• Quotes and references: open the source. Confirm the quote, and that the source really says what is claimed. Never include a reference you have not opened.
• Numbers: work them out again with a calculator or spreadsheet, and compare with the original data.
• Names, dates, titles and places: check them against a reliable source.
• Claims about law, health, money or safety: treat the answer as a starting point, and confirm with a qualified source.
• Anything you cannot check: remove it, flag it, or label it clearly as unchecked.

Ways to make checking cheaper

• Give it the source material and say use only this. Then ask it to quote the line that supports each claim, so you can check fast.
• Ask for assumptions and doubts: list what you assumed and what you are least sure about.
• Ask for the other side: now argue the strongest case that this conclusion is wrong.
• Get a second opinion. Ask the key question in a different assistant, or in a fresh chat, and compare. A disagreement shows you where to look.
• Use tools for what they do well. Use spreadsheets and code for maths, and search for recent facts.
• Check by sampling. For a long result, check five random items. If any fail, check everything.

Review your previous answer as a skeptical fact-checker.
1. List every factual claim, number, name and date in it.
2. For each, say: supported by the material I gave you / from your general knowledge / uncertain.
3. Flag anything you may have invented or that may be outdated.
4. Tell me the 3 claims I should verify first, and where I could check each.
Do not defend the answer; try to find its weaknesses.

Match the checking to the risk

• Low stakes, such as brainstorming or a personal first draft: skim it and do a sanity check.
• Medium stakes, such as anything sent to a colleague or client: check every fact and number, and read all of it.
• High stakes, such as legal, medical, financial, safety, compliance, published or customer-facing work: AI helps, a qualified person decides, and a second reviewer checks.

Protect private data

• Do not paste anything you would not be happy to have stored by an outside company, unless your plan and your policy clearly allow it.
• Before pasting, remove or replace names, account numbers, health details and secrets. Use placeholders like Client A.
• Know the settings: whether chats are used for training, how long they are kept, and whether work accounts differ from personal ones.
• Follow your employer's AI policy. If there is none, ask before you use client or company data.
• Be open about it. If AI helped a lot with something people expect to be your own work, say so.

Your personal AI-use policy

1. I never send, publish or submit AI output without reading all of it.
2. I verify every quote, number, name and date before it leaves my hands.
3. For legal, medical, financial or safety content, AI drafts - a qualified person decides.
4. I do not paste client secrets, personal data or credentials into AI tools. I use placeholders.
5. Agents get the least access needed, and I approve irreversible actions myself.
6. I note AI involvement where it matters to the reader.
7. Every quarter I review my saved memory, my playbook and these rules.