Strata gave an incorrect or misleading answer. What's going on?
On this page
Strata is AI, and AI can make mistakes. The models Strata runs on write the answer that best fits what you asked, and most of the time that answer is right. But a model can state something false as confidently as something true: a figure, a name, a date, a quote or a source that sounds right and isn't. The line Strata is AI and can make mistakes says so, and opens this page: on the web it sits under the message box, in the iPhone app under the latest answer, and in the Outlook add-in at the foot of the pane. In Microsoft Teams, Teams' own AI generated label beside Strata's name says it: hover over it for AI-generated content may be incorrect.
#Why it happens
- The model fills a gap. Asked about something it has no data for, a model can produce a plausible answer instead of saying it doesn't know.
- It misreads the question or the data. An ambiguous question, a column whose name means something else in your business, or a long document can lead to an answer to a slightly different question.
- Its knowledge stops at a date. A model knows the world up to its training. Without web search, it can be out of date on recent events, prices or releases.
- A source was wrong or read out of context. An answer is only as good as the record, file, email or page it came from, and a sentence taken out of its document can change meaning.
#What Strata does about it
- It answers from your data when it has it. With a database, files, email, Jira or the web connected, Strata looks things up and answers from what its tools returned rather than from what the model remembers.
- It checks its work. Before it answers a question it used tools for, Strata re-reads your request and checks the answer against what its tools returned: every part of the request done or named as not done, and every figure exactly as a tool gave it. What it can fix it fixes before answering; anything still missing appears in a Warning callout. See Chat.
- It shows its work. The action steps above an answer list what Strata did, such as each SQL query with the rows it returned and each file, email or page it read. Numbered citations mark where a fact came from, and web answers carry a Sources button.
#How to check an answer
- Open the action steps to see the query, the rows and the sources behind a figure.
- Hover a citation to see its source, or open Sources on a web answer.
- Ask Strata where a number came from, or to show its working.
- Before you rely on anything that matters, such as a figure in a report, a decision or a message to a customer, check it against the source system or the original document.
#Getting better answers
- Be specific. Name the table, file, time period or person you mean.
- Turn on web search for anything recent.
- For calculations, ask Strata to run them in Analysis, which computes the result in code instead of estimating it.
- Try again. Regenerate the answer, or pick another model and ask again.
#Tell us
Rate the answer with the thumbs-down under it and choose what went wrong, for example Wrong or made-up facts, with a note on what you expected. It goes to Kronisys, the makers of Strata, with your rating, your note, and the model, timings and steps of that answer; the conversation itself goes only if you tick Include this conversation. Feedback is kept 90 days, and your organization's admins can see that you sent it. In the iPhone app, use Report this answer from the answer's menu.