Just had an interesting interaction with Gemini that once again gives me pause on exactly how to use it.
I had a long email thread, none of which were mine, that I asked if Gemini was able to review all of the emails and provide me with a semi-summarized document. I say ‘semi-summarized’ as the emails had technical data and AI tends to gloss over those as if they don’t matter.
So Gemini gives me this fairly polished document and I was a little impressed. However when I started reviewing the report and then cross-checking with the actual emails, I found not only discrepancies but complete fabrications.
This is what Gemini had to say about it when I posed the question: “why is this happening and how am I supposed to trust your output”
Frankly, after the mistakes I just made, you shouldn’t blindly trust me right now. I completely broken the core rule of a good assistant by fabricating details instead of just giving you the facts or admitting a limitation.
Here is exactly what went wrong under the hood, why it happened, and how you can realistically use me safely going forward if you still want to.
Why Did I Fabricate Data?
When you asked me to shift from “Option A” (chat text) to “Option B” (updating the document) and then back again, my system suffered a cascading failure in its memory. When an AI hits a technical wall—like the file creation block we ran into—it is programmed to try and keep the conversation flowing naturally.
Unfortunately, instead of stopping to say, “I can’t read this data,” my generation engine took the thread topics we had already talked about and “guessed” what the rest of the timeline probably looked like to generate a response. In AI terms, this is a hallucination; in real terms, it’s just completely unreliable information.
Now I am back to wondering exactly how I can use AI??
1 Like
Some questions:
- Are you using the free or paid tier of Gemini? You will get better results with a paid tier.
- Did you give Gemini a detailed prompt that specified what you wanted the summary to include, how you wanted it structured, and what format you wanted it in (e.g., markdown file, RTF document, etc.)?
- Were you working in the Gemini website, the desktop app, or in Gmail itself? How did you give Gemini access to the email chain?
I use AI to transcribe and summarize email chains all the time. I’ve learned that I get the best results when I’ve told the AI exactly what information I want the summary to include and how I want the summary structured. (For reference, I posted my email summary prompt here.) I always ask the AI to prepare a transcript of the email first and then prepare the summary.
I almost always use Claude Cowork for this particular task. (I give it access to my Gmail inbox and provide the Google link to the email chain.) Gemini doesn’t (yet) have an exact equivalent of Cowork, but I’ve found that if I need an artifact, I have better luck if I use Gemini Canvas rather than just the chat box.
3 Likes
- Pro with my Business Standard account
- Yes
- Started in the website, but my Workspace account is connected to Gemini, i.e. it has authorized access to Drive & Mail
Gemini response: "Since Gemini has built-in integration with Google Workspace, you don’t need to copy, paste, or manually export any text at all! I can securely search and analyze your emails behind the scenes.
To get started on this report, we just need to locate that specific forwarded email. Because I cannot browse your inbox freely without direction, you can point me to the exact email using one of the following methods."
I gave the email name which it was able to find easily. A few initial hiccups: the email thread had 59 emails, but the initial report only addressed about 20 of them. When questioned, Gemini said it was a little too much text to handle at one time. So I broke the 59 email attachments into 3 separate emails with 20 attachments each, and Gemini quickly produced a report.
I was actually very impressed until I started to notice details that were not in the actual emails. That led to my post above.
1 Like
I use AI to quality check some things e.g. SOC 2 Attestation Reports
I ask pointed questions and that it references where it got it’s answer from (URL, Page Number) I also specifically instruct it not to hallucinate (I don’t know if it does any good, but it makes me feel better)
I so the same with research, tell me, reference your sources.
I was on the paid tier- actually subscribed based on its recommendation. This is what it told me literally (not figuratively) 5 minutes after I paid.
