You can trust AI enough to give you a first draft, a summary or an outline you are going to read and check, but not enough to send its facts, numbers or promises to anyone without checking them first. The more an answer could cost you if it is wrong, the more you check, and Catch Made-up Facts Before They Reach Customers teaches the routine in seven minutes.
AI tools write in the same confident tone whether they are right or guessing. That is called a hallucination when the guess is wrong, and it is the reason "trust" is the wrong frame. Think "check in proportion to the risk" instead.
Which tasks are low risk and which are high risk?
Low risk: trust it, then skim.
- Rewording your own text, fixing grammar, changing tone.
- Summarizing a document you have open and can check.
- Brainstorming ideas, headlines or questions to ask.
- Turning your own notes into a list or a table.
Medium risk: check every fact before you use it.
- Customer emails, where AI may add a policy or promise you never made.
- Spreadsheet analysis, where one total should be checked against your records.
- Research summaries, where a source may be old or second-hand.
High risk: a qualified person decides.
- Legal, tax, medical or HR decisions.
- Anything that changes money: refunds, prices, payroll.
- Anything sent to a client, regulator or employee in your company's name.
The rule underneath: AI can draft, a person signs off.
How do sourced tools change the picture?
Some tools show where each fact came from, which makes checking much faster.
- Gemini Notebook (formerly NotebookLM) answers only from the documents you add and cites the passage.
- Perplexity searches the web and cites its sources.
- Felo answers with a numbered source on every claim.
- ChatGPT and Claude can search the web and add citations when you ask, and Claude's Research mode writes a cited report.
A citation is not proof. It is a shortcut to the page you still need to open.
What does checking look like in a real lesson?
In Research: a cited competitor price scan, Claude's Research mode checks what lawn care companies near Bright Path Landscaping charge and writes a report with a citation on every price. You open one source and look at the small print at the bottom of the page: the prices are from March 2024, and the report did not say so. That single check is the difference between talking points you can use and an embarrassing quote.
In Felo: Price a New Service in 10 Minutes, a made-up bookkeeping firm researches what a fractional CFO package costs. Felo's table says one firm's prices are firm-published, but when you open source 40 it is a comparison article, not the firm's own website. You ask Felo where every price came from, and the new table links to the firms' own pricing pages and marks the rest as third-party.
With your own documents, Gemini Notebook lets you click a citation number and see the highlighted passage in a cleaning contract, so you can confirm the answer before you rely on it.
How do I check an answer quickly?
- Ask what it assumed. "List every fact in your answer that I did not give you."
- Open one source for each number you will act on, and read the date.
- Check one total in any analysis against your own records.
- Read customer-facing text as the customer, looking for promises.
- Tell it to use only your facts and to ask instead of guessing.
More detail: How to check AI output before you use it and How to stop ChatGPT from making things up.
What mistakes should I avoid?
- Trusting a citation without opening it. Sources can be old, second-hand or misread.
- Checking the easy parts and skipping the numbers. Numbers and dates are where errors hide.
- Assuming a confident tone means a correct answer.
- Letting AI decide refunds, prices or anything legal.
Key takeaways
- Trust AI for drafts you will check, not for facts you will send.
- Match the checking to the risk: skim, verify or escalate.
- Sourced tools speed up checking, but open the source.
- A person signs off on anything that reaches clients, money or legal decisions.
Frequently asked questions
- How often does AI make things up?
It varies by tool, task and how much context you give it, so assume any answer can contain a made-up detail. Giving it your facts and telling it not to guess reduces the risk.
- Is Perplexity more trustworthy than ChatGPT?
Perplexity shows its sources by default, which makes checking easier. Its answers still need the same checks: open the source and read the date.
- Can I use AI answers in a report to my boss?
Yes, after you have checked every fact and number and can say where each one came from.