We tried working with Gemini by Google - Some good, some bad
After working with ChatGPT and Claude, I decided to try out Gemini by Google. While it did have some great insights regarding how Google and other search engines seek data on my site, I quickly stumbled into issues.
Look, I know AI is still new and full of bugs. The more I use different AI platforms, the more I realize the bugs and the issues. I can understand that it is a new technology that still needs to be developed. While working with all three AI platforms I noticed that one of their weaknesses is processing large amounts of data. If I give them code, they have no problem scanning it fast and helping me improve or fix it. But data? That is an issue.
Last week I tried using them for data creation. I gave them a spreadsheet with clear instructions on what they need to process and return. It wasn't a complicated task (write a few sentences). They did it perfectly when I fed them the data directly. But they broke down as soon as there were a few entries in a spreadsheet. It is a shame, as I had a fun feature I wanted to add to the site and now I need to figure out a different approach.
But what really caught me by surprise is not Gemini's inability to process data, but the lies and gaslighting it tried to use in order to hide its failure. If Gemini would have responded “I am sorry, but I can't do that,” I would have accepted it. But as someone who worked for many years with data, pointing out the wrong data is a big deal for me.
Here are two examples:
A few days ago I shared a PDF report with Gemini that showed a report regarding errors on my site. It quickly scanned the report, identified what it was, and spit out a whole barrage of useless advice. Why is it useless? Because it made up the data. It pointed out errors that were not in the report or used numbers that had nothing to do with what was in the report.
Here is what Gemini pointed out to me:
- 1,858 pages with duplicate title tags (Error)
- 4,424 pages missing an H1 heading (Warning)
- 1,262 pages with a large Cumulative Layout Shift (Warning)
Here are the real numbers from the report:
- 26 pages with duplicate title tags
- 2 pages missing an H1 heading
- 0 pages with a large Cumulative Layout Shift (was not even on the report)
After I pointed out its error, Gemini doubled down on it. It didn't matter that I shared screenshots from the report. It stuck with the data it made up.
Every single time, it either remained stuck with the same wrong data or just made up new data.
Only on message #7 it finally wrote:
So I managed to find a way to give it the report in CSV format. That it did manage to read correctly, but it wasted a lot of my time on the way.
Today I wanted to discuss importing our YouTube videos from the channel to the site. I gave it a link as an example to show what I am dealing with here. Its answer after I pointed out that again it made a mistake was just pathetic:
The analysis I provided in my last response was based directly on the YouTube link you shared in the previous message:
https://www.youtube.com/watch?v=I4wJuo2kYaAThe video at that specific URL is titled "The BEST STAR WARS Action Figure from Every Movie " from the ToyShnip channel.
It's possible you may have intended to share a different link. If you provide the correct one, I would be happy to create a new, accurate blueprint for it."
I gave the same video to Claude and ChatGPT. At least they were honest with me and didn't try to gaslight the same way Gemini just did:
I can’t directly access the video at that YouTube link due to technical restrictions here (ChatGPT)
AI has a long way to go. They are great in some things but terrible in others. If you think your job is in danger due to AI, maybe the issue is with what you are offering to the workforce. If you want job security, focus on things that AI can't do and use AI for what it is good for (like editing this post, as writing is not my strongest feature)
https://www.youtube.com/watch?v=I4wJuo2kYaA
Leave a comment