A while back I made a preliminary version of my rules for using AI safely. I’ve let it sit for a long time, so it’s time to revisit it! This time I’m going to put down five rules (five? Probably five, though some are maybe sub-rules … we’re going to play this by ear) and then I’ll come back and write a blog post for each rule.

As a reminder, these are rules for an individual — not an organization or society or government or world or anything. These are just how you, an individual person, can more safely use AI.

1. Understand what AI is

AI is one of the few technologies where there’s basically zero learning curve — if you know how to converse you can sit down, chat with an LLM, and get something out of it that looks like a good response.

This means a lot of people use AI without having any idea what’s going on under the hood. This can lead to some unintended consequences — it can make people assume AI always tells the truth. They can assume AI “understands” them (they can assume AI “understands” anything).

I’m not saying you need to become a computer scientist or anything, what I’m saying is that you should understand AI about as well as you understand a car.

See, you drive a car, and there are some things about it you understand that help you drive. You know the gas goes in the gas tank. You know the engine makes that gas explode (in very small quantities) and that makes something go around. You likely know that the transmission takes that thing that’s turning around and connects it to the wheels so that they start turning, causing you to move forward.

You might know that there are discs attached to the wheels and when you step on the brakes some things squeeze the discs and that slows you down. You know that when the brake pedal goes “dugadugadugaduga” that’s the antilock brakes and your car is trying to keep you from sliding. If you don’t drive a BMW you know that the little stick by the steering wheel makes the blinkers light up.

You don’t know everything — you’re not a mechanic — but you know enough to safely use the car to get from point A to point B, and part of “enough” means a basic understanding of what goes on under the hood.

You don’t need to know everything about LLMs to use them, but you should have a basic idea of what is going on under the hood. That will save you from some very serious mistakes.

2. Double check AI outputs

There are two main reasons to double check AI outputs:

1. AIs can hallucinate — that is, they can just make stuff up. This happens more than you realize and sometimes it’s harmless, but other times it can be incredibly harmful and if you’re not paying attention it’ll slip right past you.
2. AIs can misconstrue true information — AIs don’t “understand” as we do, so it can take something out of context, repeat it to you confidently free of context, and you’ll never know

In either case the end result is the same — you are given some information that appears confidently correct, but is not. If you base decisions on it you will be wrong. If you pass it to someone else they will believe you are wrong. If you base important decisions on it you could cause serious harms.

A corollary to this is that in order to double check AI you must be capable of double checking it.

What I mean is that if you ask AI something you know nothing about, and something where you can’t read the sources with any confidence, then you have to accept whatever the AI says at face value — you’re incapable of knowing if it is outputting truth or not.

If you’re incapable of double checking AI then the AI can’t be double checked, and you are putting yourself in danger. Don’t use AI for things you are incapable of double checking it!

For example, if you ask AI to review a long, complicated legal document because you don’t understand the legal jargon, are you going to know if the AI is answering your questions correctly? Since you already don’t understand the source document the answer is no, you have no way of knowing if it’s doing a good job. And that could get you in serious legal trouble!

This is one area where people get in deep trouble with AI. Multiple users have proudly proclaimed that they have created whole new branches of physics or mathematics with their AI. Were they mathematicians or physicists before? No! So AI told them some stuff that sounded amazing to them, and then told them it was groundbreaking, and they just went with it — only to discover later when people who WERE capable of double checking AI work told them it was garbage that they were being led along by an algorithm. Always double check, and if you can’t double check, don’t do it.

3. Treat AI like the world’s most confident intern

Let’s expand on the idea of double checking with a mental model that goes even farther — treating AI like an intern.

Obviously you would double check an intern’s work, and the more important the closer you’d read it.

But you also need to be careful about what information you give that intern. For example, you might not want to give the intern the ability to add, remove or change appointments in your calendar at will. You might not want them reading all of your emails, or responding on your behalf without double checking.

More importantly, you don’t know what an intern is going to post about online or share with strangers. If someone particularly persuasive gets in touch with an intern they could bluff them into revealing your banking details or your secrets or anything else.

This is especially troubling because many AI assistants encourage you to give them LOTS of access — basically all the access. And on the one hand, the more access you give them the more they can help. The more they can do proactively.

BUT, the more access you give them the more damage they can do, too.

I work in cybersecurity and I know a lot of people who work in the field and we experiment with AIs all the time. But most of the time we isolate them, or give them very specific read-only privileges to subsets of our information. We take pains to reduce the blast radius.

If you’re not familiar with how to restrict permissions or isolate an AI agent then you’re putting your data at risk when you give an AI access to it. So don’t do that! Treat it like an intern. Give it only the information and instruction it needs to do a job, then make sure you double check the work it produces.

4. Learn the latest in using AI well

Every day there are new developments in AI — new ways to prompt, new methods of reducing hallucinations. This isn’t a post that teaches you how to use AI well — we’re talking about how to stay out of trouble.

But there are techniques that you need to learn to use it well, and those will also keep you out of trouble. There are some models that are good at some things and not others. There are some strategies that work well in some circumstances, but not in others.

Take the time to learn the skill of using AI well. I can’t really expand on it here because it’ll be different tomorrow, but one thing that will likely always be true is that some organizations will put out regularly updated “how-tos” about the latest techniques for using AI. Find those and stay up to date.

5. Don’t use AI for something you need or want to be good at

What was the heaviest thing you lifted today? And how much did it weigh?

If you have children, it was probably one of them. Maybe you brought in the groceries. Maybe you were doing some yardwork.

For most people, the heaviest thing they lift in any given day is, well, weights. They work out, they do strength training, whatever.

They do this because our bodies need that kind of activity to stay healthy, but we created a world full of machines that lift and move things for us, so that we don’t really need to do any heavy lifting on a given day.

So, we created machines that lifted heavy stuff for us, and that leads to us being less healthy, so we created machines that allow us to lift heavy stuff.

It’s the CIIIIIIRRRRRRCCCCCLLLLEEE OF LIIIIIIFFFFEEEEEEE

What we’re talking about is “biomechanical offloading” (there’s actually several terms, but that’s what we’re using). It’s the act of using something else so we don’t have to do a physical activity. When we experience biomechanical offloading our body will do less activity and will thus start to atrophy — it will become weaker.

To fix that, we lift weights. I guess that’s biomechanical onloading.

When you use AI to do something you are performing cognitive offloading. That is, you’re taking something you typically did with your mind, and you’re having something else do it.

This isn’t something AI introduced. In fact, you can probably think of a few ways you’ve already experienced cognitive offloading (depending on your age):

• You might have learned how to do math in your head at school, but once you got a calculator on your smartphone those skills probably disappeared
• When you were younger you may have gotten around with a map, or just by listening to someone tell you directions. But since the advent of GPS in cars you probably use it anywhere you’re going somewhere new — and maybe even when you’re doing a trip you are already familiar with
• Using search engines has probably caused you to remember fewer facts (this is the “google effect”) — can you remember when the Nicene Creed was written? Probably not, but you immediately thought “Well, I could just google it” — even if you memorized that date in school
  (I did memorize that date in school, and forgot it, and googled it — it was 325 CE)

Not all cognitive offloading is bad, but any cognitive offloading does reduce the associated skill. For some things that’s not the end of the world (nobody is going to quiz you on the Nicene Creed unless you’re on Jeopardy), but for other things it can be a serious problem. And AI is especially dangerous because it can do SO MUCH, and it’s so easy.

One example I like to use is chatting with whatever gender you’re interested in. If you’re single, and you want to mingle, you might be tempted to chat with someone and ask AI to sharpen up your wit or just help you come across as more educated than you are. This could work via text, but what about when you end up across from them on a date? What will you do then?

YOU WILL CRASH AND BURN AND NEVER FIND LOVE!!!! I’M SORRY!

But that is the truth. If you are relying on ChatGPT to communicate with people then when the time comes to communicate in person you will not be the same person they have been communicating with, and that is REALLY off-putting.

The only way to get good at conversation is to converse. The only way to get good at professional communication at work is to communicate professionally at work. If you have AI write your rough draft you will never gain the skills you need, or you will slowly lose them. Either way, your ability to communicate will atrophy.

The same holds true for dozens of other cognitive skills. If you want to be creative, but you’re always asking AI to brainstorm for you, you will lose what creativity you have. If you want to be a great leader, but ask AI to punch up your team meetings, you will never be able to rally the troops in a crisis.

I could go on but I think I’ve made my point. If you want or need a skill, DO NOT hand it over to AI. You will experience cognitive offloading, but there’s no Gym for the brain (yet) — no way to artificially build the cognitive muscles that are atrophying.

If you want or need to be good at something, don’t use AI.

That’s all for now

Those are the five rules I’ve put together so far (some earlier rules were combined into these). I feel pretty good about them! But there are still some major blind spots so I’ll likely update this as time goes on.

The most important thing, the unwritten sixth rule, if you will, is to pay attention to how your life changes when using AI. Not all changes are for the best, and it’s up to you to pay attention to your life and make sure it’s going in the right direction.

You know what? That’s an incredibly important rule, and there’s still more that need to be added or merged into existing rules later. I’m going to throw a reminder list here of other rules I may add — sound off if any sound particularly interesting:

1. Pay attention to how your life changes (definitely going to add next edit of this page)
2. The risk of workslop (or personalslop, I guess) — less but better vs. more but average
3. Addiction, Emotional dependence and isolation (don’t let AI replace the people in your life)

To be continued!