New to AI? Start here → Start 1. ChatGPT vs. Codex: what's the difference?
Check AI answers and results
In one line
Turn a convincing answer into a checked result
Evidence over tone, real behavior over “done” messages.
The big picture
Illustration (may look a little different from the real screen)
Follow along
A confident tone doesn’t mean it’s correct
AI can write the wrong date or number, or point to sources that don’t exist. The more important it is, the more you should treat the answer as a draft and check it against sources and real results.
Separate facts from opinions → open the sources → compare with the original → review before you use it
Open source links and check them yourself
- Does the page actually exist?
- Is the source qualified to make this claim?
- Do the publish date, update date, and the time it applies to match?
- Does it actually back up the claim you’re checking?
Even with an official website address attached, the link may have nothing to do with the claim. “It has a source” and “I checked the source” are different things.
Recalculate numbers, starting with a small sample
For a total, add up a few rows yourself and check the units and time period. For a percentage, ask what it’s a percentage of. For a work report, compare against the numbers in the original file before using it.
Check the periods being compared, the old value, the new value, and how the growth rate was calculated.
With Codex, check the results and the changed files
“Task complete” on its own isn’t enough. Look at what changed and what checks were done. For a web page, open it yourself and try typing, the buttons, and different screen sizes.
Illustration (may look a little different from the real screen)
Tell me which files changed, what you actually checked works, and what you haven't checked yet.
Sample replyChanged file: index.html
Checked: code syntax
Not checked: typing and saving in a real browser
Next check: add a book, then refresh
Tell me which files changed, what you actually checked works, and what you haven't checked yet.
Sample replyChanged file: index.html
Checked: code syntax
Not checked: typing and saving in a real browser
Next check: add a book, then refresh
In this example only the code syntax was checked, so you can’t conclude that saving works. Reading what was and wasn’t checked is the key.
Have a person review before anything important
Before sending an email outside, deleting files, publishing a document, or changing payment or account settings, check the content first. When you first automate something, starting with made-up data and copies makes it easier to compare results.
For high-stakes topics like health, law, and money, don’t decide based on AI answers alone. Check with trusted sources and experts in that field.
Prompts you can copy
Change the parts in square brackets to fit your situation, then paste it into ChatGPT or Codex.
Check what an answer is based on
Review a Codex result
Where people get stuck
Quick check: Codex said it only checked the code syntax. Can I assume saving works too?
Answer: “No, check saving separately”
Checking syntax and checking that it actually works are different. You need to try the real flow yourself: typing, saving, and refreshing.
Official sources
I'm stuck here
Stuck while following along? Tell me where. Only the site owner sees this. Leave your email only if you'd like a reply.
💌 Get new tips by email
I'll email you when a new tip goes up. No ads, ever.