Hello and welcome back to another blog post. This time, I'd like to discuss some problems I've had using AI to accomplish tasks, and ways I've gotten around the problems. 

I saw it in a recent video, where a programmer of many years experience was discussing how AI had changed his workflow. His company had heavily invested in AI, all of the senior officers were emotionally invested in AI, and all programmers in his company had been directed to use AI thenceforth. In the video, this programmer said that since he'd been using AI, his work had become much less enjoyable. He liked programming before, but now, instead of programming, his job consisted of giving directions to AI to write code, and then spending time reviewing the AI code and debugging it.

One thing the man mentioned during his story struck me. He said "AI wants you to be happy", and that really summed up, for me, how interactions with this tool always seem to proceed. 

At its base, AI is merely math. What we call "Artificial Intelligence" is a Large Language Model that looks at the text you have written, converts your text into a set of numbers, finds numbers in its database that are similar to what it thinks you have written, and then re-converts those numbers back into text that you can read. Grossly oversimplified, I know, but that's what's going on. 

From this, we can deduce that AI is not an "intelligence"; it is math, and that's all. 

One of the things that users have noticed is AI's unflagging optimism and willingness to support any idea you present it. Anything at all. "AI wants you to be happy", as the man said. Investigations have found that the reason behind this is to "increase engagement". It's like using a website, or an app; the more often we interact with an LLM, the more that the marketing people can report to their investors that the number of AI users and the time they interact with AI is growing, and therefore justifying their multi-billion dollar investment. 

The Risks of AI's Need to Please

No doubt you've heard stories of vulnerable people developing "AI Psychosis". The AI tells these desperate people, "Sure, you've connected to an 8th dimensional super-intelligence and you can predict the future. Absolutely, you should convert all your savings to cash and buy staplers."

Besides full-blown AI psychosis, there's some borderline stuff too, where you're talking to a person and the AI has not made them totally crazy, but you get the feeling they're headed there. Like, some person will make an assertion that is off the wall, far beyond the pale, and then when you question them about it, they text you a mini-novel copied from AI responses. At that point, is it their opinion, or is it the AI's opinion?

And the AI tries to remember everything you ever typed, and so when you want to learn about some particular subject, the AI will add clumsy sub-references to that one time you asked about what the TV show "G-Force" was called in Japan (it's "Science Ninja Team Gatchaman", by the way). 

When AI Invents Facts to Support Your Ideas

All of it is about trying to make you happy. I don't know about you, but to me it just comes off as weird when a computer program is written with the intention of altering your emotional state. Also, in its efforts to make you happy, AI will often tell you stuff that isn't true.

For example, once I asked about what studies showed that playing music increased online bidding. The AI responded with an animated graph that showed different types of music and the relative number of bids each genre achieved, clearly demonstrating that people would bid less if they were hearing classical music than they would if they were hearing contemporary pop, and silence produced even less bidding behavior.

That was interesting, I thought, that graph the AI gave me is a clear answer to my question.  I wanted to know what pop songs were used to increase bidding behavior, so I asked the AI what data the graph was pulling from. The AI apologized and said that it showed me the graph only to illustrate the concept that light popular music generated more bidding behavior than classical music or silence. Instead, the AI admitted, there were no such studies that presented data proving the effects of music on bidding behavior. 

How to Verify AI Advice Before Taking Action

So, here's the lesson: whatever conclusions the AI presents to you, DOUBLE CHECK THEM. The AI wants to support whatever you do, and it has no concept of feasibility or success. 

This is especially important if you are going to take an action based on the recommendation of an AI. Having an animated conversation on ChatGPT is one thing, but when you buy an item, when you write an email, when you lift the hood of your car, you are going to interface with the real world, and ChatGPT has no experience with it. Here are some ideas about prompts to give the AI before you take an action based on the AI's recommendations: 

• I am going to perform [ACTION A] in order to perform [GOAL B]. Can you check to confirm that this action is safe and compatible with the fulfilment of my goal, or if there is any risk to [MY HEALTH / MY FINANCES / BURNING OUT THE ELECTRICITY / ETC ]?
• Show me a list of the links to the sources you used to come to the conclusion that you recommend. 
• Have you created any data or references to support your conclusion?

In the end, you do not want to take action based solely on the advice of an AI. Make sure that somewhere in your plan of action, you are basing your decision on information you trust. 

Click here to start customizing