Bots have one weakness that worries people more than any other: they sometimes state things that aren't true, with full confidence. See Hallucinations.
ISO has a principle that speaks right to it.
Evidence-based decision making
One of ISO's seven quality management principles is evidence-based decision making: decisions based on the analysis of data and information are more likely to give the results you want (iso.org).
ISO's principles now go further, asking organizations to recognize the limits and possible bias of data, analytics and automated tools. That's a standards body, in plain words, reminding everyone that automated tools need checking too.
The bot rule
For bots, the principle turns into one rule:
Never invent facts. Trace every number to a dated source.
And two helpers:
- If you can't confirm it, leave it out, or say clearly that it's unconfirmed.
- Say when the data is partial. "This covers 40 of the 90 files" is honest. "Here's the answer" is not, if it's only part of the data.
What a source looks like
| Weak | Strong |
|---|---|
| "Studies show…" | A link to the study |
| "About 50%" | "54 of 90, from the template review on Oct 21" |
| "The price is $12" | "The price is $12 (price list, rev. Oct 2026)" |
| "Everyone does this" | "Most of the 90 templates did this" |
The strong version can be checked. That's the whole point.
Why this works on bots
A model fills gaps because it's built to give a likely-looking answer. Telling it "it's okay to say you don't know" removes the pressure. Telling it "every number needs a source" gives it a habit to follow. Neither is perfect. Both help a lot.
Then you check. See Verification and evals.
A real example
Our AN Creation bots share a core principle: never invent facts; trace every spec to its source. In our review of 90 published marketplace templates, 54 had some version of the "never invent facts" rule, but only 10 asked the bot to say when data was partial (good-practice.md).
This site runs on the same rule. Product facts are checked against xAI's docs. Numbers carry their source. Things we can't confirm are left out, or presented as our view.
Don't fake precision either
A related trap: precise-sounding numbers with no basis. "This saves 37% of your time." If nobody measured it, it's made up, even if it sounds scientific. Round, honest words ("in our use, it saved a lot of retyping") beat fake numbers.
What this means for you
- Put "never invent facts" in every bot's principles.
- Ask for a source and date on every number that matters.
- Make "I couldn't confirm that" a safe answer.
- Click a source now and then. Make sure it says what the bot said it says.
Evidence is what turns a confident answer into a trustworthy one.