What Does AI Know About You?

The genesis of the question

Every once in a while, I will read some articles saying hiring managers are looking for AI skills, even for positions in accounting or marketing. Or maybe something along the lines of hiring managers will not hire anyone without AI skills.

Those types of articles leave me frustrated because I am left with the question of what kind of specific AI skills are hiring managers looking for? Is it machine learning? Is it LLM training? Prompt engineering? Or maybe something simpler such as filling in a box with a question?

If they are looking for something along the lines of machine learning, then that will be a lot more years of upskilling. But if it is something simpler such as being comfortable with using CoPilot or ChatGPT, that is a whole different ball game.

And another question comes up that kind of worries me: how do you generate value from CoPilot or ChatGPT if hiring managers are talking about the simpler version of AI skills? I’ve been experimenting with generative AI and so far, I am not seeing major innovation out of them. Yes, you can have it summarize your emails or your Team meetings and maybe produce PowerPoint, but for some of us, especially in finance, those are not where the real value lies. For the finance types, value might be spotting potential fraud, introducing new ideas to make money based on consumer sentiment, analysis of future trends, etc.

That’s the kind of stuff I’m looking for and right now, the biggest block is the risk of showing confidential data to those generative AI. There are significant privacy issues to think of.

AI discussion group formed

I decided to introduce a new discussion forum on AI for those who were interested. I wanted to hear what others know about AI and how they were using it. So far as I can tell, most aren’t really using generative AI in any fascinating way that would add value to businesses. If they were using it, it was mainly for summarization or to write emails, things of that nature.

The other thing I wanted to do was do some kind of AI project – a simple one. Nothing too complicated and out of our skillset. Most of us are not techies.

In one of our group sessions, someone told a story about a friend that asked generative AI where her location was. The AI told her it didn’t know; it was not privy to such information. Then she asked, “where’s the nearest McDonald?” and the AI gave the correct location.

It actually knew her location.

That drove another person to try that experiment and he found that it was able to give him his correct location. However, he was puzzled by its ability because he was using VPN (which should be hiding his location) so he asked it how it knew his location. It responded with “…based upon some conversations we have had…” (These are not exact quotes.)

These initial queries led to a session where we would test the AI to learn what it knew about us. I am not going to detail the exact path of our queries but instead will focus on high level findings out of this one session, lasting one and a half hour.

The experiment was mostly with CoPilot since it was readily available and seemed easier to query and have conversation. Someone tried Brave but Brave did not allow one to have a conversation like CoPilot.

Initial findings from experiment

• Generative AI at this time will not divulge information from LinkedIn – at least not that night. We made quite a few attempts to pull work information but to no avail. This restriction might be just a CoPilot problem.
• However, if you have any information sitting out in the web, it will find it.
• I had done an experiment in ChatGPT where I had uploaded a document detailing stories about my work accomplishments. I wanted to check to see if ChatGPT would leak to CoPilot any of my wonderful achievements. From the admittedly brief tests, it appears that ChatGPT may be firewalling that information.
• We did delve into somebody’s ancestors to see if CoPilot could find those particular websites or YouTube. The results were mixed. The more common the name is, the harder it is to locate the correct information.

Closing summary

The findings are necessarily limited due to the time limit of one hour and a half. I do find the fact that CoPilot did not query LinkedIn very interesting.

For the most part, Microsoft appears to have put in place fairly strong guardrails to prevent personal information from being provided. The stuff out in the web, though, might hold personal information. I wish governments would go after those people and jail them because putting out personal information is like doxxing.

Enough for tonight. I have more stories about our explorations to tell you but that will be for another day.