Common ways AI misleads
- Invented figures. A model can state a number that is not in any source. Always check numbers against the filing.
- Stale data. An answer can be confident and out of date. Check the "as of" date.
- Confident tone. Fluent prose can hide uncertainty. Ask what the tool could not verify.
- Backtests that look too good. A strategy fitted to past data can look strong and still fail. Ask whether results are out of sample.
A safe workflow
- Ask a specific question, such as "What did the company say about margins in its last announcement?"
- Check the source the answer cites, and confirm the date.
- Ask for the counterargument: what would make this view wrong?
- Decide yourself, using your own rules for position size and risk.
- Record the decision and why, so you can check later whether the research held up.
Questions to ask any AI tool
- Where does each figure come from, and when was it produced?
- What does the tool say when data is missing?
- Does it explain how each figure is produced?
- Does it place trades or manage money?
- Who checks its outputs before you see them?
Where Atradian fits
Atradian is designed to show sources and dates, to mark what it cannot verify as unknown, and to show data freshness and known gaps on the reliability page. It does not place trades. You still make the decision, and the methodology is set out on the methodology page.