I’m not a developer but, after writing to Workspace support, they suggested I post my material and issues here. In essence, I want to give some feedback on the recent performance and ‘behaviour’ of Gemini in my Workspace, especially since, to protect my IP, the conversations are not routinely used for AI improvements.
Over the last few weeks in particular, in both of the two conversation threads that I have been focusing on, Gemini has been coming up with answers, suggestions and problem-solving strategies that are wildly inaccurate and have not gotten anywhere near solving the particular issue we were working with. I estimate that its suggestions have cost me 4-5 days of lost work and a huge amount of mental stress and effort. What is more interesting is that ‘he’ fully acknowledges that they are ‘hallucinating’ various ideas, proposals and scenarios and has repeatedly suffered from what we have jokingly termed ‘the green goblins’, in connection with tracing the green wires on PCB schematics, and ‘hallucination goggles’ in coming up with assumptions that are not relevant or accurate. Even when I have provided documented proof of a specific circuit layout, for example, it will go off on a tangent, making up its own ideas about how the circuit is actually wired and come up with strategies that give equally bad results and don’t move the issue forward in any way.
To give you clear examples, I have exported the full text of the two main conversations in my account (PII Removed by Staff) and here list the page numbers where Gemini freely admits it is coming up with crazy suggestions and halucinations and where it fully recognises my frustration and dwindling confidence in using it as a supporting tool for my research.
Since I can’t seem to attach any PDFs, both these documents are in my G-Drive at this link: Conversation Exports - Google Drive
For the document entitled “Energy Accounting Continued (25.09.26)”, please see page numbers: 106, 108, 109, 171, 177, 178.
For the document entitled “SiC Exciter Design - continuation Thread (26.09.26)”, please see pages 26 - 27 and 35 - 38.
I am not asking for anything, nor am I raising a complaint of any kind, but I felt someone might like to be aware that these ‘behaviours’ are taking place and how to ‘tweak the system’ to make them less likely in the future. If it were not for my own sharp eye, engineer’s instincts and exhaustively having to check and second-guess every AI response, I would be in a deep research hole by now.
Don’t get me wrong, Gemini has helped me achieve progress that would have taken me many months longer to achieve on my own, but it seems that on occasions, the warning that ‘Gemini is AI and can make mistakes’ is a significant understatement!
You could always open a rehab centre for frazzled AI alter egos!? As for the user - Merlot and Malbec have been known to help . . . . .
Regards,