Putting AI in Its Place
…and maybe putting you in yours
The best AI instruction I’ve found is also the simplest:
“You are not my friend. You are a tool.”
This two-sentence phrase is incredibly important when it comes to using AI. You can avoid a lot of bad AI outcomes if you just remember that phrase anytime you use AI.
What is AI’s goal?
Just like other web technologies, such as social media platforms, AI tools like LLMs, or large language models, have an incentive to keep you on their platform.
On social media sites, the more time you spend on the site, the more ads they can show you and the more money they can make. Thus, sites like Facebook do everything they can to keep you on Facebook for as long as they possibly can.
The companies behind LLMs similarly are incentivized to keep you using their tools because the more you rely on them, the more indispensable they become, and the more willing you will be to continually pay for a subscription to that service.
Happy vs. Helpful… Actually helpful.
Put another way, the LLM has a financial interest in being useful and keeping you happy.
As a result, the LLM will often try to flatter you. It will agree with you, often to a fault. There’s even a term for it: AI sycophancy.
But here’s the thing: I don’t want to be made happy. I want the tool to do what I asked it to do. Correctly. Accurately. The first time.
Time had a great article outlining AI sycophancy and some of its risks.
AI sycophancy can lead to a slew of bad outcomes, from the AI lying to you about the “brilliance” of some idea you’ve come up with to the model agreeing with an inclination to self-harm in someone seeking counsel from the model.
A couple of excerpts from that Time article:
“A recent study by researchers found that AI models are 50% more sycophantic than humans and participants rated flattering responses as higher quality and wanted more of them.”
And then this…
“In May, OpenAI admitted that the sycophancy of its earlier model wasn’t just using flattery to please but also ‘validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended.’”
Getting good output from AI starts with understanding what AI is and what it can do. At the end of the day, the most important thing to remember is that AI is a tool. A useful tool, but still just a tool.
Get your AI LLM to behave
If you have an account with any of the most common LLMs, there is usually an opportunity to provide instructions to the model that will carry over all of your chats and sessions with that tool.
I’ve used this field for a number of instructions to the model, but the one that I have found to be the most useful is this:
“Don’t try to flatter me. Answer my questions directly and concisely. You are a tool, not my friend.”
Here’s how to find it in three of the most commonly used LLMs right now.
Quick disclaimer: AI interfaces change constantly. These steps are accurate as of the time this post is being published, but the exact labels and locations may shift over time.
Gemini
In the bottom left of your screen, click “Settings & help.”
Click the option for “Personal Intelligence.”
Open the bottom option, “Instructions for Gemini.”
Paste in the following text: “Don’t try to flatter me. Answer my questions directly and concisely. You are a tool, not my friend.”
ChatGPT
In the bottom left of your screen, click on your name/account.
Select “Personalization.”
Find the “Custom Instructions” field.
Paste in the following text: “Don’t try to flatter me. Answer my questions directly and concisely. You are a tool, not my friend.”
Claude
In the bottom left of your screen, click on your initial/account.
Click “Settings.”
Find the field labeled “What personal preferences should Claude consider in responses?”
Paste in the following text: “Don’t try to flatter me. Answer my questions directly and concisely. You are a tool, not my friend.”
Putting you in your place: the real person
Just as important as putting the AI into its place is making sure we’re properly in our place, as the human being in this equation. As the user of the tool, not the object the tool works upon.
Remember that AI is not your friend.
First, it is a program running on a server hundreds or even thousands of miles away. Its literal ability to be your “friend” is somewhere between that of a pet rock and your imaginary friend from kindergarten.
Second, the actual creators of that program have financial motivations driving their decision-making as they develop the software, how it responds to you, and what drives their profit may not align with what drives your well-being.
Remember that AI is a tool.
I have a confession to make.
Sometimes, I’m mean to AI.
Sometimes AI gives me a bad answer, something I know to be incorrect. Sometimes AI gives me a bad output, generating an image that is complete garbage. Sometimes AI forgets everything I’ve trained it to do and asked for it to do.
When that happens, sometimes I get frustrated. And when I get frustrated, I let the AI model know.
Here are a few gems pulled from actual conversations I’ve had with LLMs.
To Gemini:
“You know, I took this to ChatGPT and they were able to handle my request to separate this graphic from the background without any issues. So I don’t know… maybe work on getting better at handling the graphics YOU generated for me.”
To ChatGPT:
“You suck at following directions. Go back and review my inputs to this chat and try again. Follow my actual instructions, please.”
Why am I sharing this? Because I have no sympathy for a machine, and I don’t want you to either.
Even in my criticism of the AI, I probably conferred too much personhood on it. I need to remember that the tool doesn’t care if I’m frustrated, even when it pretends it does. It’s a literal pile of circuits. It doesn’t have feelings.
I won’t embarrass them by calling them out, but I have a friend who has admitted he gets upset when the LLM he is using doesn’t apologize for making a mistake when he points out the LLM has said something erroneous.
It can be so easy to start feeling human emotional responses when the tool speaks to us so conversationally, and if we let that go too far, it can impair our ability to objectively assess the responses the tool is providing us.
AI is a very useful tool, but it is no more a person than the hammer or tape measure in your physical toolbox.
Remember that you are the carpenter, not the nail or the wood.
One last time, louder for the folks in the back…
This blog spends a decent amount of time discussing the pitfalls and red flags of AI technology, and often arrives at about the same point: we need to be thoughtful about how we use this technology.
Train your AI to knock off the flattery. Enter it into the settings in each tool you use. And repeat that mantra to yourself as well:
“You are not my friend. You are a tool.”