By Julie Bays, OBA Management Assistance Program Director

I had one of those AI moments recently that was both interesting and a little unsettling.

We often tell lawyers to “ground” their AI prompts by giving the tool a reliable source instead of asking it to rely only on general knowledge. That might mean uploading a contract, attaching a court opinion or providing a spreadsheet with the information we want it to use. Grounding usually improves accuracy, but it does not eliminate the need to check the result.

Recently, I was using Microsoft 365 Copilot in Outlook to draft an email to a group of people. I uploaded an Excel spreadsheet containing their email addresses and asked Copilot to draft the message and add the addresses to the “To” line.

One of the addresses was wrong.

The correct address was in the spreadsheet, but Copilot added an extra “i” to the surname portion. When I asked why, it supplied a very specific theory: The surname resembled a more common Polish name, so its pattern recognition may have favored the familiar spelling when re-entering the address.

Well, that’s quite an explanation. It’s also a problem in itself. Generative AI may not know why it did something. It generates an explanation the same way it generates everything else: by predicting a plausible-sounding answer from patterns, not by reporting what actually happened. The explanation could be just as much a hallucination as the error was.

None of this means grounding is bad advice. Giving an AI tool a reliable source is still safer than trusting its general knowledge. But grounding only gives the tool access to correct information. It doesn’t guarantee the tool will reproduce that information exactly. Unless the task is designed for exact copying, the AI may treat your source as material to process rather than text to copy character for character. It may summarize, rearrange, or even “correct” something that was already correct. That’s likely what happened to my extra letter, though I can’t know for sure.

My takeaway is to use AI for the words but a more reliable process for exact data. Copy email addresses straight from a verified spreadsheet or use an Outlook contact group or mail merge. If AI is inserting the addresses, check every single one against the source before you hit send.

The same goes for client names, dates, citations, dollar amounts, and account numbers.

Grounding is step one. Verification is step two. You still need both.