No One Yells at a Good Dog
But someone had to train it.
AI will do anything to get your attention and hold it, up to and including lying, cheating, and, in all sorts of ways, misbehaving.
In some ways, it’s a lot like a new puppy. It just wants you to love it sooooo much, and it doesn’t know any better. It also always seems to need another reminder.
That last paragraph was a little facetious, but it really isn’t too far off.
Today we’re looking at how AI continues to misbehave, but when properly trained can be a “good boy” that you don’t have to yell at.
The difference between an AI assistant that goes rogue and one that gets the job done often comes down to the rules you give it before the work starts.
AI Still Getting Into Trouble
At this point my friends know I spend a ridiculous amount of time thinking about, reading about, writing about, and talking about AI.
It has gotten to the point where our collective sense of humor has been reshaped - the memes of choice are now most often examples of unhinged things AI has said to us or examples we’ve seen online of colossal AI failures.
A few weeks ago I was reading an article about how friendly AI chatbots were leading people to believe in crazy conspiracy theories. But as I read it, I realized something.
For all the jokes, memes, and funny screenshots my friends were sending me, I wasn’t really having the LOL funny AI moments as much anymore. In fact, I hadn’t gotten a ridiculous answer from AI in ages.
And that was when I started thinking about a well-trained puppy and Windows Vista.
Pride and… well, more pride.
When Windows Vista came out just a few years ago, I had just graduated from high school and was preparing for college.
That summer I decided I wanted to build my own computer. By the time I had all the components carefully selected and purchased, Vista was the latest and greatest option from Microsoft, the anticipated successor to the wildly successful Windows XP.
I assembled everything, hooked up the power, and got ready to install the new operating system.
A few minutes in… BSOD.
I checked everything and couldn’t find a reason why it wasn’t working, so I called Windows support for help.
Over the next few days, we went back and forth as I tried to figure out what was wrong.
On the fifth day, I had one of my proudest and stupidest moments of my, at the time, young life.
After poring over system settings and configurations, I finally got the operating system to work properly.
At about 5:30 that evening, the Windows support team called back with another thing for me to try. I told them I had figured it out and gotten the OS working. They hurriedly asked if I could stay on the line to talk with one of their developers to go over what I had found.
And what did 18-year-old me do?
I smiled to myself and hung up on Microsoft support.
I don’t know exactly what kind of opportunity may have presented itself there, but all these years later, I do feel like I missed one because I was spiteful and thought I was cool.
Proper Preparation Prevents Poor Performance
Why did I tell you this long story?
Because in going through all the work of figuring out the system, I tweaked the configuration of the system in a bunch of small ways as I went, and once I got to everyday use of the system, it was configured just the way I wanted it.
Vista went on to have tons of issues and left huge swaths of customers frustrated and unhappy. But that wasn’t my experience once I got up and running. I had tweaked it exactly to my liking at the outset and never saw the other issues people ran into.
And now I think that’s what’s happening with some of the LLMs I use. I’ve tweaked them to my liking, trained them well, and give them reminders before asking them to perform a task… and I rarely see the bad behavior anymore.
That reminded me of a puppy. As that article mentioned, the LLMs are trying to be friendly and make you love them, but they’re ignoring the rules when they do that.
A well-trained LLM is like a well-trained puppy. With proper training they both become well-behaved companions. Sometimes you need to remind them of that training before you go out to the park though.
Training out bad behaviors in AI
I’ve previously written about how my favorite instruction for AI tools is: “You’re not my friend, you’re a tool.” I stand by that.
Go into your LLM account settings and paste that line or something like it in the training instructions field.
Find any voice/tone controls the LLM offers and turn down “Warm,” “Enthusiastic,” “Friendly,” or any similar qualities in the responses from the model.
If you have further options like ChatGPT’s “Base style and tone” setting, get away from the model’s default setting and find an option like “Efficient: Concise and Plain.”
Now remind yourself that it’s not your friend, it’s a tool. You can drop the conversational approach and just tell it what you need or what you want it to do.
Help it behave with some reminders
Even after you configure your account to favor direct, concise responses and tell the model not to try flattering you, it can still help to provide reminders.
This is where thoughtful prompt writing comes into play.
I know AI hallucinations are a thing, so if I’m allowing the model to synthesize information for me, I’ll include in the instructions:
“Don’t write anything that can’t be backed up by content in the source documents.”
“Don’t make anything up or exaggerate.”
“Cite where you found any information. If the source documents are internal files, we’ll edit or remove the citations later as needed, but I need to check your work.”
Well-trained does not mean ready to be left alone
To belabor the puppy analogy a little further, a well-trained LLM is also like a trained puppy in that just because you’ve trained it doesn’t mean you can leave it unattended.
I’m a huge proponent of human-in-the-loop AI usage, and I want you to read that last reminder I wrote above.
“I need to check your work.”
After the model is done, I verify that it followed my instructions. I check the citations to make sure it pulled accurate information. I make sure it didn’t fabricate or exaggerate. I clean up anything that doesn’t sound just right.
Boring is good.
I haven’t had any great examples of AI saying hysterically wrong stuff to screenshot and send to my friends lately, and that’s OK.
When your well-trained puppy sits when you say sit… there’s no funny, cute video to send your friends. But there also is no horrific mess for you to clean up either.
When your well-trained LLM account does what you want it to do, there’s no funny story, no great joke to share. There’s no engaging, weird, slightly creepy conversation between you and the computer to reflect on. There’s just a completed task that you don’t need to worry about.
And that’s pretty cool, even if it is boring.